JP3689616B2 - Voice recognition apparatus, voice recognition method, voice recognition system, and program recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention provides speech recognition with a speaker normalization function that is used in the learning mode / recognition mode by transforming (normalizing) the spectrum as if a standard person uttered the speech uttered by a certain speaker. The present invention relates to a device, a speech recognition method, a speech recognition system, and a program recording medium in which a speaker normalized speech recognition processing program is recorded.
[0002]
[Prior art]
  So far, a technique for speaker adaptation or speaker normalization with a relatively small amount of speech data has been developed in a speech recognition apparatus constructed by learning a large amount of speech data in advance. In addition, it is desirable that the speech data required by the speech recognition apparatus for speaker adaptation and speaker normalization be as small as possible in consideration of the utterance burden of the provider.
[0003]
  As a method of speaker normalization with very little data, speaker normalization is performed using the lengths of the vocal tract divided into two sections, the front (oral side) and the rear (pharyngeal cavity side). There are methods to perform (for example, Naito, Li Deng, Sakazaka, “Speaker normalization method using speech generation model”, Acoustical Society of Japan Proceedings, pp.99-100 (March 1999)).
[0004]
  On the other hand, there is a report that speaker non-speaker speech recognition performance has been improved by performing simultaneous nonlinear expansion and contraction of the frequency axis and intensity axis of the speech spectrum (for example, Nakagawa and Kamiya). , Sakai, “Recognition of word speech of unspecified speakers based on simultaneous nonlinear expansion and contraction of time axis, frequency axis and intensity axis of speech spectrum”, IEICE Transactions, Vol.j64-D No.2, pp.116 -123 (1981)).
[0005]
  Furthermore, DP (dynamic programming) matching in the frequency domain was performed between spectral envelope sequences (n-dimensional vector sequences) of a plurality of vowels in the speech uttered by the conversion source speaker and the conversion destination speaker in advance. A method for converting the spectrum envelope of the conversion source speaker into the spectrum envelope of the conversion destination speaker using two optimal DP paths has been proposed (for example, Japanese Patent Laid-Open No. 4-147300).
[0006]
[Problems to be solved by the invention]
  However, the conventional speaker normalization method has the following problems. That is, in the speaker normalization method using the length of the front (oral side) and back (pharyngeal cavity side) of the vocal tract, the body such as the vocal tract shape and the length of the vocal tract is among the individual differences of the speakers. Individual differences (hard differences) due to the above differences can be normalized to some extent, but individual differences (soft differences) due to vocalization such as articulation points and mouth openness cannot be normalized. Therefore, there is a problem that the recognition performance is not sufficiently improved by normalization.
[0007]
  In addition, in the method of speaker normalization by simultaneous non-linear expansion / contraction of the frequency axis and intensity axis of the above-mentioned speech spectrum, not only individual differences but also phoneme differences can be normalized unless the constraint conditions for non-linear expansion / contraction are set fairly well. As a result, there is a problem that the performance is lowered.
[0008]
  Further, in the method of performing DP matching between the spectrum envelope sequences (n-dimensional vector sequences) of a plurality of vowels in the speech uttered by the conversion source speaker and the conversion destination speaker, the utterances such as the articulation point and the degree of opening of the mouth If the optimal DP path for each vowel is different due to the individual difference (soft difference) caused by, the other group is biased toward those with more members of the similar optimal DP path group (for example, back tongue vowel). However, there is a problem that a slightly inappropriate DP path is extracted and a DP path that is not optimal as a whole is selected. In addition, when learning vowels can be successfully selected so that the optimal DP path is not biased, only individual differences (hard differences) due to physical differences such as vocal tract shape and vocal tract length are normalized. Since it is a DP path, there is a problem that the recognition performance is not sufficiently improved by normalization. Furthermore, since the conversion source speaker and the conversion destination speaker are based on the restriction that the same content (word or sentence: for example, “Aiueo, No Aio”) is spoken, the content of the speech of the conversion source speaker is different. There is also a problem that it cannot be used when there is a lack of audio data.
[0009]
  Thus, the above conventional speaker normalization method is not sufficient in terms of the normalization performance of individual differences.
[0010]
  Therefore, an object of the present invention is to normalize not only individual differences (hard differences) caused by physical differences but also individual differences (soft differences) caused by vocalizations with a small amount of voice data and normalizing phoneme differences. Disclosed is a speech recognition apparatus and speech recognition method, a speech recognition system, and a program recording medium on which a speaker normalized speech recognition processing program is recorded.
[0011]
[Means for Solving the Problems]
  In order to achieve the above object, a first invention is a speech recognition apparatus with a speaker normalizing function for converting a first speaker's voice into a second speaker's voice and normalizing the speaker, A spectral envelope extraction means for extracting a first spectral envelope from the first voice uttered by the first speaker and extracting a second spectral envelope from the second voice uttered by the second speaker; A first memory means for storing a spectral envelope and a second spectral envelope with a label of a speech unit; and for the same label, the first spectral envelope and the second spectral envelope stored in the first memory. A nonlinear frequency axis spectrum matching means for performing frequency nonlinear matching using dynamic programming to obtain a frequency warping function representing a correspondence between the frequency axes of both spectrum envelopes; The second memory means for storing the warping function with the label of the speech unit and the first spectrum envelope of the recognition target extracted by the spectrum envelope extraction means are inputted, and the speech unit of the recognition target first spectrum envelope is inputted. A frequency envelope conversion means for reading out a frequency warping function having the same voice unit name from the second memory and converting the first spectrum envelope to be recognized into a spectrum envelope relating to the second speaker based on the read frequency warping function And speech recognition means for recognizing input speech based on the converted normalized spectral envelopeA frequency warping function stored in the second memory means based on the label. , Similar phonemes , Voiced interval / Group by silent section and speaker , Calculate the average value of the frequency warping functions belonging to each group , The spectrum envelope transforming means comprises the recognition target stored in the first memory, comprising an averaging means for assigning a label of each group name to the obtained average frequency warping function and storing the average frequency warping function in the second memory means. The average frequency warping function of the group determined according to the amount of data of the first speaker related to the same phoneme as the phoneme of the first spectrum envelope is used as the frequency warping function.It is characterized by that.
[0012]
  According to the above configuration, the frequency warping function representing the correspondence between the frequency axes of the first spectrum envelope obtained from the voice of the first speaker and the second spectrum envelope obtained from the voice of the second speaker is used. The frequency axis of the first spectrum envelope to be recognized by the first speaker is nonlinearly expanded and contracted to be converted into the spectrum envelope by the second speaker, and the input speech is recognized based on the converted normalized spectrum envelope. . Therefore, if the first speaker is an unspecified speaker and the second speaker is a standard speaker, the spectrum envelope based on the utterance by the unspecified speaker is speaker normalized to the spectrum envelope of the standard speaker, Based on this normalized spectral envelope, highly accurate speech recognition is performed.
[0013]
  In that case, by obtaining the frequency warping function based on an average value of a plurality of first spectrum envelopes and an average value of a plurality of second spectrum envelopes for the same voice unit, individual differences (hard It is possible to normalize not only differences but also individual differences (soft differences) caused by vocalizations. In such a case, the phoneme difference is not normalized because it is nonlinearly expanded and contracted between the same phonemes.
[0014]
  Furthermore, the average frequency warping function is “phoneme”. , "Similar phonemes" , `` Voice interval / It is required for each group for each "voiceless section" and "speaker". Therefore, an appropriate group average frequency warping function can be selected and used in place of the frequency warping function according to the amount of utterance data of the first speaker stored in the first memory means. For example, back tongue vowels / o / If there is little or no utterance data, / o / Back tongue vowels that are similar phonemes / a / Average frequency warping function or an average frequency warping function of a voiced sound section is selected. In this way, it is possible to cope with a case where the amount of utterance data of the first speaker is small. Furthermore, by obtaining the average frequency warping function for each phoneme, individual differences due to vocal utterances are normalized.
[0015]
  In the speech recognition apparatus according to the first aspect of the invention, the first memory means stores the slopes of the first spectrum envelope and the second spectrum envelope with labels of speech units, and stores them. The first spectrum envelope slope is extracted from the first voice uttered by the first speaker, while the second spectrum envelope slope is extracted from the second voice uttered by the second speaker and stored in the first memory means. And reading the slope of the first spectrum envelope and the slope of the second spectrum envelope having the same speech unit name as the speech unit of the first spectrum envelope to be recognized from the first memory, and calculating the difference between the two slopes. On the basis of the above, a spectrum inclination correcting means for correcting the inclination of the normalized spectrum envelope obtained by the spectrum envelope converting means is provided, and the speech recognition means is adapted to correct the inclination. It is desirable to recognize the input speech based on the normalized spectral envelope.
[0016]
  According to the above configuration, the normalized spectrum envelope is based on the difference between the slopes of the first and second spectrum envelopes in the same speech unit name as the speech unit of the first spectrum envelope to be recognized by the first speaker. Is corrected, and the speaker is normalized to the second speaker's voice more accurately.The
[0017]
  MaIn the speech recognition apparatus according to the first invention, the speech unit is a phoneme, and the slopes of the first spectrum envelope and the second spectrum envelope stored in the first memory means are phoneme based on the labels.・ Group similar phonemes / voiced / unvoiced intervals and speakers, calculate the average value of the slope of the spectral envelope belonging to each group, and label the resulting average spectrum slope for each speaker name and each group name Is added to the first memory means and stored in the first memory means, and the spectrum inclination correcting means is configured to determine the average spectrum inclination of any group to which the phoneme of the recognition target first spectrum envelope belongs to the spectrum envelope. It is desirable to use it as an inclination.
[0018]
  According to the above configuration, the average spectrum inclination is obtained for each group of “phonemes”, “similar phonemes”, “voiced / unvoiced sections”, and “speakers”. Therefore, according to the amount of the first speaker's utterance data stored in the first memory means, an appropriate average spectral inclination of the group can be selected and used instead of the spectral envelope inclination. In this way, it is possible to cope with a case where the amount of utterance data of the first speaker is small. Furthermore, by obtaining the average spectral slope for each phoneme, individual differences due to vocalizations are normalized.
[0019]
  The speech recognition apparatus according to the first invention isRecordingThe spectrum inclination correcting means is used according to the amount of data of the first speaker related to the same phoneme as the phoneme of the first spectrum envelope to be recognized stored in the first memory.RuheiDesirably, a group of average spectral slopes is determined.
[0020]
  According to the above configuration, the averaged speech is averaged within the optimum range according to the amount of utterance data by the first speaker related to the recognition target phoneme.TabiraA high recognition rate can be obtained even when the amount of utterance data of the first speaker is small by using the average spectral gradient.
[0021]
  In the speech recognition apparatus according to the first aspect of the invention, when the speech recognition means obtains the frequency warping function, the extracted time series of the first spectrum envelope or the second spectrum envelope is recognized by the unspecified speaker speech. It is preferable that the recognition is performed by the method and the speech unit name of the recognition result is transmitted to the first memory means.
[0022]
  According to the above configuration, the voice unit name for label is automatically obtained from the first and second spectrum envelopes extracted from the utterances of the first and second speakers. In this way, the labeling process for the spectrum envelope or the slope of the spectrum envelope is easily performed.
[0023]
  The speech recognition apparatus according to the first aspect of the invention includes a waveform analysis means for extracting prosody information by analyzing the waveforms of the voices of the first speaker and the second speaker when there are a plurality of the first speakers. In the learning mode for obtaining the frequency warping function, the prosodic information and spectrum envelope extracted from the speech of multiple known speakers are added as a speech standard pattern with each speaker's label, while the speech of an unspecified speaker is registered. Recognition mode, the recognition target speaker is recognized using the registered speech standard pattern based on the prosodic information and spectrum envelope extracted from the speech of the recognition target speaker, and the recognition result speaker is recognized. Speaker recognition means for outputting a name, and the spectrum conversion means and the spectrum inclination correction means are configured to cause the frequency word relating to the speaker name from the speaker recognition means to be in the recognition mode. It is desirable to read out the mapping function or the first spectral slope.
[0024]
  According to the above configuration, in the recognition mode, the speaker is recognized, and the frequency warping function or the first spectral inclination related to the recognized speaker is read and used for normalization of the recognition target first spectrum envelope. . Therefore, when there are a plurality of speakers as the first speaker and the frequency warping function, the first spectrum envelope and the first spectrum slope are stored for each first speaker, the unspecified story is stored. Speaker normalization speech recognition is performed on the speaker.
[0025]
  Further, the speech recognition apparatus according to the first invention uses the speech recognition means as a non-specific speaker speech recognition method using a standard pattern extracted from the normalized spectrum envelope obtained by the spectrum envelope conversion means. It is desirable to perform voice recognition by using
[0026]
  According to the above configuration, the speech recognition unit performs speech recognition with high discrimination accuracy using the standard pattern normalized by the speaker. Further, when the speech recognition means is a speech recognition means using a mixed distribution HMM (Hidden Markov Model), the variance of the phoneme standard pattern for each speech unit is reduced, and the number of distributions used for the mixed distribution of transition probabilities is “1”. "Or" less number ". Therefore, the memory capacity and processor load during execution are reduced.
[0027]
  The second invention is a speech recognition apparatus for converting a voice with the voice quality of the first speaker into a voice with the voice quality of the second speaker and normalizing the speaker, wherein the first speaker speaks. A vocal tract cross-sectional area extracting means for extracting a second vocal tract cross-sectional area from a second voice uttered by a second speaker while extracting a first vocal tract cross-sectional area from the first voice First memory means for storing a vocal tract cross-sectional area and a second vocal tract cross-sectional area with a label of a voice unit, and the first vocal tract cross-sectional area and second stored in the first memory with respect to the same label Nonlinear vocal tract to obtain vocal tract warping function representing the correspondence of vocal tract axis of both vocal tract cross sections by performing nonlinear vocal tract expansion and contraction using dynamic programming for vocal tract cross section Axis matching means and second memory means for storing the vocal tract axis warping function with a label in units of speech The first vocal tract cross-sectional area of the recognition target extracted by the vocal tract cross-sectional area extraction means is input, and the vocal tract axis warping function having the same voice unit name as the speech unit of the recognition target first vocal tract cross-sectional area is obtained. A vocal tract cross-sectional area conversion means for reading from the second memory and converting the recognition target first vocal tract cross-sectional area into a vocal tract cross-sectional area related to a second speaker based on the read vocal tract axis warping function; Speech recognition means for recognizing input speech based on the converted normalized vocal tract cross-sectional areaAnd a vocal tract axis warping function stored in the second memory means based on the label , Similar phonemes , Voiced interval / Group by silent section and speaker , Calculate the average value of the vocal tract warping function belonging to each group , Averaging means for assigning a label of each group name to the obtained average vocal tract axis warping function and storing it in the second memory meansWithThe vocal tract cross-sectional area conversion means is a group determined according to the amount of data of the first speaker relating to the same phoneme as the phoneme of the recognition-target first vocal tract cross-sectional area stored in the first memory. The average vocal tract warping function is used as the above vocal tract warping functionIt is characterized by that.
[0028]
  According to the above configuration, the vocal tract axis warping function is used, and the vocal tract axis of the first vocal tract cross-sectional area to be recognized by the first speaker is nonlinearly expanded and contracted to be converted into the vocal tract cross-sectional area by the second speaker. Then, the input speech is recognized based on the converted normalized vocal tract cross-sectional area. Therefore, if the first speaker is an unspecified speaker and the second speaker is a standard speaker, the vocal tract cross-sectional area based on the utterance by the unspecified speaker is the speaker of the standard speaker. Normalized, and highly accurate speech recognition is performed based on the normalized vocal tract cross-sectional area.
[0029]
  In that case, by obtaining the vocal tract axis warping function based on the average value of the plurality of first vocal tract cross-sectional areas and the average value of the plurality of second vocal tract cross-sectional areas for the same voice unit, It is possible to normalize not only individual differences caused by (hard differences) but also individual differences caused by vocalizations (soft differences). In such a case, the phoneme difference is not normalized because it is nonlinearly expanded and contracted between the same phonemes.
[0030]
  Furthermore, the above average vocal tract axis warping function is “phoneme”. , "Similar phonemes" , `` Voice interval / It is required for each group for each "voiceless section" and "speaker". Therefore, an appropriate group average vocal tract warping function is selected and used instead of the vocal tract warping function in accordance with the amount of utterance data of the first speaker stored in the first memory means. Can do. For example, back tongue vowels / o / If there is little or no utterance data, / o / Back tongue vowels that are similar phonemes / a / The average vocal tract axis warping function or the average vocal tract axis warping function of the voiced sound section is selected. In this way, it is possible to cope with a case where the amount of utterance data of the first speaker is small. Further, by obtaining the average vocal tract axis warping function for each phoneme, individual differences due to vocal utterances are normalized.
[0031]
  A third invention is a speech recognition method having a speaker normalizing function for converting a first speaker's voice into a second speaker's voice and normalizing the speaker, wherein the first speaker Extracting the first spectrum envelope from the first voice uttered while extracting the second spectrum envelope from the second voice uttered by the second speaker; and for the same voice unit name, the extracted first Performing non-linear frequency expansion and contraction using dynamic programming on the spectrum envelope and the second spectrum envelope to obtain a frequency warping function representing the correspondence between the frequency axes of both spectrum envelopes; Based on the step of extracting one spectrum envelope and the frequency warping function having the same speech unit name as the speech unit of the extracted recognition target first spectrum envelope, the first spectrum envelope of the recognition target is calculated. Converting to a spectral envelope for the second speaker, and recognizing input speech based on the converted normalized spectral envelopeAnd the frequency warping function based on the label , Similar phonemes , Voiced interval / Group by silent section and speaker , Calculate the average value of the frequency warping functions belonging to each group , Assign a label for each group name to the resulting average frequency warping functionWithThe step of converting the recognition target first spectral envelope into the spectral envelope related to the second speaker depends on the amount of the extracted first speaker data related to the same phoneme as the phoneme of the recognition target first spectral envelope. The average frequency warping function of the group determined in this way is used as the frequency warping function.It is characterized by that.
[0032]
  According to the above configuration, as in the case of the first invention, the frequency warping function is obtained based on an average value of a plurality of first spectrum envelopes and an average value of a plurality of second spectrum envelopes for the same voice unit. Thus, not only individual differences caused by physical differences but also individual differences caused by vocalizations can be normalized. In such a case, the phoneme difference is not normalized because it is nonlinearly expanded and contracted between the same phonemes.
[0033]
  Furthermore, the average frequency warping function is “phoneme”. , "Similar phonemes" , `` Voice interval / It is required for each group for each "voiceless section" and "speaker". Therefore, an appropriate group average frequency warping function can be selected and used in place of the frequency warping function according to the amount of the utterance data of the first speaker stored. For example, back tongue vowels / o / If there is little or no utterance data, / o / Back tongue vowels that are similar phonemes / a / Average frequency warping function or an average frequency warping function of a voiced sound section is selected. In this way, it is possible to cope with a case where the amount of utterance data of the first speaker is small. Furthermore, by obtaining the average frequency warping function for each phoneme, individual differences due to vocal utterances are normalized.
[0034]
  The speech recognition method of the third invention extracts the slope of the first spectral envelope from the first speech uttered by the first speaker, while the second speech from the second speech uttered by the second speaker. Based on the step of extracting the slope of the spectrum envelope and the difference between the slope of the first spectrum envelope and the slope of the second spectrum envelope of the same speech unit name as the speech unit of the recognition target first spectrum envelope It is desirable to provide a step of correcting the slope of the normalized spectrum envelope, and to use the normalized spectrum envelope after the slope correction in the step of recognizing the input speech.
[0035]
  According to the above configuration, the normalized spectrum envelope slope is corrected in the same manner as in the first aspect of the invention, and speaker normalization is more accurately performed to the voice of the second speaker.
[0036]
  The fourth invention is a voice recognition system including a user terminal capable of exchanging voice information via a communication network and a center system, wherein the user terminal includes the first invention or the second invention. The speech recognition apparatus of the invention and transmission control means for transmitting the normalized spectrum envelope or the normalized vocal tract cross-sectional area relating to the input speech obtained by the speech recognition apparatus to the center system via the communication network On the other hand, the center system includes a control unit that receives the normalized spectral envelope or the normalized vocal tract cross-sectional area transmitted from the user terminal via the communication network, and the normalized Recognize input speech to the speech recognizer based on spectral envelope or normalized vocal tract cross-sectional area It is characterized by having a force speech recognition unit.
[0037]
  According to the above configuration, speech recognition with high individual difference normalization performance is performed as in the case of the first invention or the second invention. Further, the normalized spectrum envelope or the normalized vocal tract cross-sectional area related to the input speech obtained by the speech recognition device of the user terminal is transmitted to the center system via the communication network, and the input speech of the center system is transmitted. The recognition unit recognizes the input voice to the user terminal. Thus, by recognizing the input voice on the center system side, the subsequent language processing is also performed on the center system side, and the configuration of the user terminal is simplified.
[0038]
  According to a fifth aspect of the present invention, there is provided a program recording medium comprising: a computer, the spectral envelope extracting means, the nonlinear frequency axis spectral matching means, the spectral envelope conversion means, the voice recognition means,Averaging means ,A speaker normalization speech recognition processing program that functions as a spectrum inclination extraction means, a spectrum inclination correction means, and a speaker recognition means is recorded.
[0039]
  According to the above configuration, the frequency axis of the first spectrum envelope to be recognized by the first speaker is nonlinearly expanded and contracted using the frequency warping function and converted into the spectrum envelope of the second speaker. Further, the slope of the converted normalized spectrum envelope is corrected based on the difference between the slopes of the first and second spectrum envelopes. Then, the input speech is recognized based on the normalized spectrum envelope after the inclination correction obtained in this way. Therefore, not only individual differences caused by physical differences but also individual differences caused by vocalizations can be normalized, and speech recognition with high performance of normalizing individual differences is performed.
[0040]
DETAILED DESCRIPTION OF THE INVENTION
  Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. In the following description, the speech unit is referred to as “phoneme”, but the present invention is not limited to this.
[0041]
  <First embodiment>
  FIG. 1 is a block diagram of a speech recognition apparatus with a speaker normalizing function according to the present embodiment. The waveform analysis unit 1 analyzes the waveform of the input speech and outputs prosodic information (power, pitch frequency, etc.) and a cepstrum analysis result. The spectrum envelope / tilt extraction unit 2 inputs the cepstrum analysis result obtained by the waveform analysis unit 1, and the spectrum envelope as shown in FIGS. 3 (b) and 3 (d), and FIGS. 3 (a) and 3 (c). The slope of the spectrum envelope (hereinafter referred to as the spectrum slope) as shown in FIG. The speech recognition unit 3 receives the prosodic information and other parameters from the waveform analysis unit 1, recognizes the speech with reference to the phoneme standard pattern stored in the phoneme standard pattern memory 4, and outputs the speech recognition result.
[0042]
  The speaker recognition unit 5 receives the prosodic information from the waveform analysis unit 1 and the spectrum envelope and the time series of the spectrum tilt from the spectrum envelope / gradient extraction unit 2. In the learning mode, the input time series of the learning target speaker T's spectrum envelope, spectrum inclination, and prosodic information is stored in the internal memory or the like together with the speaker number as the speaker's speaker voice standard pattern. On the other hand, in the recognition mode, the time series of the input spectrum envelope, spectrum inclination, and prosody information of the recognition target speaker T is compared with the speaker speech standard pattern group accumulated in the learning mode, and the most likely Recognize a high speaker and output a speaker number as a recognition result.
[0043]
  The matching / average processing unit 6 averages the spectrum envelope and the spectrum tilt from the spectrum envelope / tilt extraction unit 2 and averages a frequency warping function described later. Further, the frequency warping function is generated by performing nonlinear frequency axis spectrum matching between the average spectrum envelopes. The spectral envelope conversion unit 7 converts the spectral envelope of the recognition target speaker extracted by the spectral envelope / slope extraction unit 2 into the spectral envelope of the standard speaker using the average frequency warping function from the matching / average processing unit 6. Then, the obtained deformed spectrum envelope is sent to the spectrum tilt conversion unit 8. The spectrum inclination conversion unit 8 corrects the inclination of the deformed spectrum envelope from the spectrum envelope conversion unit 7 using the average spectrum inclination from the matching / average processing unit 6, and recognizes the obtained normalized spectrum envelope as a voice. Send to part 3. As a result, the speech recognition unit 3 can obtain a speaker-recognized speech recognition result.
[0044]
  FIG. 2 is a detailed block diagram of the matching / averaging processor 6. The spectrum envelope / tilt memory 11 stores the spectrum envelope and the spectrum tilt extracted by the spectrum envelope / tilt extraction unit 2 with a phoneme label as a speech recognition result for each speaker by the speech recognition unit 3. The nonlinear frequency axis spectrum matching unit 12 performs nonlinear frequency matching between the average spectrum envelope of the standard speaker S stored in the spectrum envelope / tilt memory 11 and the average spectrum envelope of the learning target speaker T. Then, a frequency warping function corresponding to the obtained optimum DP path is obtained, and a phoneme name is given and stored in the frequency warping function memory 13.
[0045]
  The averaging processing unit 14 reads out the spectral envelope and spectral inclination of each phoneme stored for each speaker stored in the spectral envelope / gradient memory 11 and the frequency warping function stored in the frequency warping function memory 13, and linearly reads them. An average value is calculated by classifying into phonemes, similar phonemes, voiced / unvoiced sections, and entire speech sections (speakers) by conversion or the like. Then, the obtained average spectrum envelope and average spectrum inclination are assigned to the spectrum envelope / gradient memory 11 by giving a label of the corresponding phoneme name, similar phoneme name, voiced / unvoiced section or entire speech section (speaker). Store. Further, the obtained average frequency warping function is given a label of the corresponding phoneme name, similar phoneme name, voiced / unvoiced sound section or entire speech section (speaker) and stored in the frequency warping function memory 13.
[0046]
  As described above, the matching / averaging processing unit 6 generates a frequency warping function for correcting excessive normalization caused by individual differences among the speakers, and sets parameters according to the degree of learning regarding the spectrum inclination. Select and average.
[0047]
  4 to 7 are flowcharts of speaker normalized speech recognition processing operations performed by the speech recognition apparatus having the above-described configuration. Hereinafter, the operation of the speech recognition apparatus will be described with reference to FIGS.
[0048]
  [In learning mode]
  4 and 5 are flowcharts in the learning mode in the speaker normalized speech recognition processing operation. First, the processing operation in the learning mode will be described in detail with reference to FIGS.
[0049]
  In step S1, an initial value “1” is set to the speaker number s. The speaker number s, the subsequent phoneme number x, the phoneme interval number y, the learning / recognition target speaker (hereinafter simply referred to as the target speaker) number sT, the standard speaker number sS, etc. are stored in the working memory ( (Not shown). Further, as the speaker, a speaker who can be the standard speaker S and the target speaker T at the time of normalization is selected. In step S2, a speech waveform is input to the waveform analysis unit 1.
[0050]
  In step S3, the waveform analysis unit 1 performs cepstrum analysis on the input speech waveform, and extracts the cepstrum analysis result and recognition prosodic information (phoneme duration, average power, average pitch frequency, etc.). The In step S4, the spectrum envelope / gradient extraction unit 2 extracts the spectrum envelope based on the low-order cepstrum coefficients from the waveform analysis unit 1. Further, the inclination of the approximate line when the spectrum envelope is approximated by a least square approximation line is extracted as the spectrum inclination. In this way, for example, if the horizontal axis is the frequency f and the vertical axis is the spectrum slope Rx (f) or the spectrum envelope Tx (f) of the phoneme x of the phoneme number x of the standard speaker S and the target speaker T, By normalization, conversion as shown in FIG. 3 becomes possible.
[0051]
  In step S5, the speech recognition unit 5 performs speech recognition using the HMM based on the time series of the spectrum envelope extracted by the spectrum envelope / gradient extraction unit 2 and the prosodic information extracted by the waveform analysis unit 1. Done. Then, a phoneme number (phoneme name) series as a recognition result is output. Here, it is assumed that the phoneme number is determined in advance in association with a phoneme name and stored in a RAM (Random Access Memory) (not shown).
[0052]
  In the present embodiment, the speech waveform analysis performed by the waveform analysis unit 1 is a cepstrum analysis, and a spectrum envelope and a spectrum inclination are extracted based on the cepstrum analysis result. However, the speech waveform analysis method in the waveform analysis unit 1 is not limited to this, and any speech waveform analysis method may be used as long as it can extract a spectrum envelope such as LPC (Linear Prediction Analysis). .
[0053]
  In step S6, the spectrum envelope and the spectrum inclination extracted by the spectrum envelope / inclination extraction unit 2 are given a label consisting of a pair of the speaker number s and the phoneme number x from the speech recognition unit 3, and matching / Stored by the spectrum envelope / tilt memory 11 of the averaging processing unit 6. In this case, in the case of supervised learning and the utterance content is known, a phoneme label string corresponding to the utterance content input to the spectrum envelope / tilt memory 11 is used instead of the phoneme number x as the speech recognition result. Use.
[0054]
  In step S7, the time series of the prosodic information extracted by the waveform analysis unit 1 by the speaker recognition unit 5 and the time series of the spectral envelope and the spectral inclination extracted by the spectrum envelope / gradient extraction unit 2 are: A label consisting of the speaker number s is assigned and stored in the internal memory as the speaker's voice standard pattern. In step S8, it is determined whether or not there is a learning voice that is uttered by the speaker having the speaker number s, that is, whether or not there is a voice input by the same speaker. As a result, if there is, the process returns to step S2, and the process shifts to extraction of the spectral envelope and spectrum inclination relating to the next voice, voice recognition, and accumulation of a voice standard pattern. On the other hand, if not, the process proceeds to step S9.
[0055]
  In step S9, the phoneme number x is set to an initial value “1”. In step S10, the averaging processor 14 reads out the spectrum envelope and the spectrum tilt to which the speaker number s and the phoneme number x are assigned from the spectrum envelope / gradient memory 11. Then, each of the read spectrum envelope and spectrum tilt is classified into “phonemes”, “similar phonemes”, “voiced / unvoiced sections”, and “entire speech sections”. In step S11, the phoneme number x is the maximum value x._MAXIt is determined whether or not this is the case. As a result, the maximum value x_MAXIf so, the process proceeds to step S13, and if not, the process proceeds to step S12. In step S12, after the phoneme number x is incremented, the process returns to step S10 to shift to the classification for the spectrum envelope and the spectrum tilt of the next phoneme.
[0056]
  In step S13, for each of “phonemes”, “similar phonemes”, “voiced / unvoiced intervals”, and “whole speech intervals” related to the spectrum envelope and the spectrum inclination given the speaker number s by the averaging processor 14 Is calculated by linear transformation or the like. Then, the obtained average spectrum envelope and average spectrum inclination are stored in the spectrum envelope / inclination memory 11 with labels of the corresponding phoneme name, similar phoneme name, voiced / unvoiced sound section, and entire speech section.
[0057]
  In step S14, the speaker number s is changed to the maximum value s._MAXIt is determined whether or not this is the case. As a result, the maximum value s_MAXIf so, the process proceeds to step S16; otherwise, the process proceeds to step S15. In step S15, the speaker number s is incremented. After that, the process proceeds to step S2, and the next speaker is shifted to the extraction of the spectral envelope and the spectral inclination, the phoneme recognition, the accumulation of the speech standard pattern, the classification of the spectral envelope and the spectral inclination, and the average value calculation. In step S14, the speaker number s is the maximum value s._MAXWhen it is determined that the above is true, the process proceeds to step S16.
[0058]
  In this way, the spectrum envelope and the spectrum inclination extracted from the speech data of the standard speaker S and the learning target speaker T are accumulated with the labels of the speaker number s and the phoneme number x. Also, the average spectrum envelope and average spectrum slope for each of “phonemes”, “similar phonemes”, “voiced / unvoiced segments” and “speech segments” are the speaker number s, phoneme name, similar phoneme name, voiced sound. Labels of the section / unvoiced sound section and the entire voice section are given and accumulated.
[0059]
  In step S16, the learning target speaker number designated from the outside is set to the learning target speaker number sT. In step S17, the standard speaker number designated from the outside is set as the standard speaker number sS. In step S18, an initial value “1” is set to the phoneme number x. Hereinafter, only the matching / average processing unit 6 performs processing.
[0060]
  In step S19, the non-linear frequency axis spectrum matching unit 12 searches the spectrum envelope / gradient memory 11 for an average spectrum envelope to which the speaker number s corresponding to the learning target speaker number sT and the phoneme number x are assigned. Is done. Based on the search result, it is determined whether or not the phoneme data for the learning target speaker is stored in the spectrum envelope / tilt memory 11. As a result, if stored, the process proceeds to step S20, and if not, the process proceeds to step S24. In step S20, the non-linear frequency axis spectrum matching unit 12 searches the spectrum envelope / gradient memory 11 for an average spectrum envelope to which the speaker number s corresponding to the standard speaker number sS and the phoneme number x are assigned. . Then, based on the search result, it is determined whether or not the phoneme data for the standard speaker is stored in the spectrum envelope / tilt memory 11. As a result, if it is stored, the process proceeds to step S21. If not stored, the process proceeds to step S24.
[0061]
  In step S21, the non-linear frequency axis spectrum matching unit 12 uses non-linear frequency axis spectrum matching based on dynamic programming to calculate the average spectrum envelope of the standard speaker S and the average spectrum envelope of the target speaker T with respect to the phoneme. Matching is done. Then, a frequency warping function corresponding to the optimum DP path is obtained.
[0062]
  FIG. 8A shows the concept of nonlinear frequency axis spectrum matching by dynamic programming executed by the nonlinear frequency axis spectrum matching unit 12. With respect to the average spectral envelope S of the standard speaker S and the average spectral envelope T of the learning target speaker T with respect to the same phoneme, the spectral envelope is equally divided into L by band, and the output values (spectrums) of both spectral envelopes S and T An element value representing (strength) is defined as an element value Si and an element value Tj (1 ≦ i, j ≦ L). Then, the frequency axis is nonlinearly expanded and contracted by dynamic programming so that both spectral envelopes correspond to each other. That is, a series of lattice points c = (i, j) on a plane composed of two spectral envelopes S and T to be corresponded
                  F = c₁, c₂, ..., c_K, ..., c_L
think of. A sequence Fmin that minimizes the sum D along the sequence F of the distance d (i, j) = d (c) between the element value Si and the element value Tj regarding the lattice point c = (i, j) The optimal DP path (frequency warping function) is used.
[0063]
  In step S22, the non-linear frequency axis spectrum matching unit 12 sends the frequency warping function together with the phoneme number x to the frequency warping function memory 13. Then, the phonetic number x and the speaker numbers sT and sS are assigned and stored by the frequency warping function memory 13.
[0064]
  As shown in FIG. 8B, the data format of the frequency warping function used in this embodiment is such that the element value of the lattice point c (i, j) on the DP path is an integer larger than “0”. The element value of the grid point c (i, j) other than the path is a matrix of L rows and L columns such that the element value is “0”. A larger number of band divisions L is preferable because warping accuracy is improved. However, if the number is too large, the storage capacity of the frequency warping function memory 13 becomes large and the processing time becomes long.
[0065]
  In the above description, the non-linear frequency axis spectrum matching unit 12 uses the element value (spectrum intensity) of each channel in the average spectrum envelope S of the standard speaker S and the average spectrum envelope T of the target speaker T for the same phoneme. Although matching is performed using Si and Tj, the matching target is not limited to the output value (spectrum intensity) of each channel of the spectrum envelope. For example, matching may be performed using the difference between output values (spectral local slopes) ΔS and ΔT between adjacent channels regarding the average spectrum envelope S and the average spectrum envelope T.
            However, ΔTj = Tj−T (j−1)
                  ΔSi = Si−S (i−1)
              Where 2 ≦ i, j ≦ L
[0066]
  In step S23, the phoneme number x is the maximum value x._MAXIt is determined whether or not this is the case. As a result, the maximum value x_MAXIf so, the process proceeds to step S25; otherwise, the process proceeds to step S24. In step S24, the phoneme number x is incremented. After that, the process returns to step S19, and the process proceeds to the process of matching the spectral envelopes of the standard speaker S and the target speaker T of the next phoneme and storing the obtained frequency warping function.
[0067]
  In step S25, the averaging processor 14 reads out the frequency warping function for each speaker from the frequency warping function memory 13, and classifies the “similar phonemes”, “voiced sound sections / unvoiced sound sections” classified in step S11. ”And“ the entire voice section ”are calculated by linear transformation or the like. Then, the obtained average frequency warping function (which may be replaced by the added value of the frequency warping function as shown in FIG. 8 (c)) corresponds to the corresponding similar phoneme name, voiced / unvoiced sound section, and entire speech section. And is stored by the frequency warping function memory 13. After that, the processing operation in the learning mode is terminated.
[0068]
  As described above, in the learning mode, the spectrum envelope and the spectrum inclination are extracted for each phoneme from the utterances of the learning target speaker T and the standard speaker S. Then, an average spectrum envelope and an average spectrum inclination are obtained for each of “phonemes”, “similar phonemes”, “voiced / unvoiced sections”, and “entire speech section”. Furthermore, a frequency warping function corresponding to the optimal DP path is obtained by performing non-linear frequency axis spectrum matching between the average spectral envelope of the standard speaker S and the average spectral envelope of the target speaker T for the same phoneme. The average frequency warping function is calculated for each of “phonemes”, “voiced / unvoiced intervals” and “entire speech interval”.
[0069]
  Thus, a speaker normalization function (average frequency warping function) for normalizing the average spectrum envelope of the learning target speaker T to the average spectrum envelope of the standard speaker S is learned. In this case, the frequency warping function is obtained using an average spectral envelope averaged by the same speaker and the same phoneme. Therefore, individual differences that differ from phoneme to phoneme caused by vocalization are normalized.
[0070]
  [In recognition mode]
  6 and 7 are flowcharts in the recognition mode in the speaker normalized speech recognition processing operation. Hereinafter, the processing operation in the recognition mode will be described in detail with reference to FIGS. Here, the spectrum envelope / gradient memory 11 of the matching / averaging processor 6 stores “phonemes”, “similar phonemes”, “voiced / unvoiced sections” and “entire voice sections” obtained in the learning mode. Are stored for each speaker as learning data. The frequency warping function memory 13 stores, as learning data, average frequency warping functions for “phonemes”, “similar phonemes”, “voiced / unvoiced intervals”, and “entire speech intervals” obtained in the learning mode. Has been.
[0071]
  In steps S31 to S35, in the same manner as in steps S1 to S5 in the flowchart at the time of learning, initial value setting of phoneme section number y, input of speech waveform, extraction of cepstrum analysis results and prosodic information, spectrum Envelope and spectral tilt extraction and speech recognition are performed. Then, a phoneme number (phoneme name) series as a recognition result is output.
[0072]
  In step S36, the speaker recognition unit 5 stores the time series of the prosodic information from the waveform analysis unit 1 and the time series of the spectrum envelope and the spectrum tilt from the spectrum envelope / gradient extraction unit 2 in the internal memory. It is compared with the speech standard patterns (prosodic information, spectral envelope and spectral tilt) groups. Then, the speaker added to the most likely speech standard pattern is recognized, and the speaker number sT as the recognition result is output. In step S37, the time series of the spectrum envelope and the spectrum tilt extracted by the spectrum envelope / slope extractor 2 is a pair of the phoneme number x from the speech recognizer 3 and the speaker number sT from the speaker recognizer 5. And is stored in the spectrum envelope / tilt memory 11 together with the section position (starting end, end).
[0073]
  In step S38, the phoneme section number y is incremented. In step S39, it is determined whether or not there is a continuous voice input to the waveform analysis unit 1. As a result, if there is an input speech waveform, the process returns to step S32 to shift to extraction of the spectral envelope and spectral tilt for the next speech, speech recognition and speaker recognition. On the other hand, if there is no input waveform, the process proceeds to step S40. In step S40, the maximum value y of the phoneme section number y_MAXIs set to the value (y-1).
[0074]
  In step S41, an initial value “1” is set to the phoneme section number y. In step S42, the spectral envelope conversion unit 7 loads from the spectral envelope / gradient memory 11 the time series of the spectral envelope and spectral gradient by the speaker sT related to the phoneme x in the yth section extracted in step S34. The In step S43, it is determined whether or not there is sufficient learning data regarding the phoneme x of the speaker sT in the spectrum envelope / tilt memory 11. If the result is sufficient, the process proceeds to step S44, and if not, the process proceeds to step S46.
[0075]
  In step S44, there is sufficient learning data regarding the phoneme x of the speaker T to be recognized (corresponding to the speaker number sT of the speaker recognition result), so the spectrum envelope conversion unit 7 causes the frequency warping function memory 13 to An average frequency warping function between the standard speaker S and the recognition target speaker T regarding the phoneme x in the y-th section is read out. Then, the spectrum envelope T by the recognition target speaker T of the speaker number sT loaded in step S42 is expressed by the following equation using the average frequency warping function (element value c (i, j)).
          Si = ΣTj * c (i, j) / Σc (i, j)
      However, 1 ≦ j ≦ L (or i−α ≦ j ≦ i + α, α: positive integer)
Thus, a modified spectrum envelope S (i-channel element value Si) of the recognition target speaker T is obtained. As a result, as shown in FIG. 9A, the channel position (j = 4) of the peak Ta of the spectrum envelope T of the recognition target speaker T is warped to the channel position (i = 3) in the modified spectrum envelope S. It is done.
[0076]
  In step S45, the average spectral inclination and the speaker number of the standard speaker S to which the standard speaker number sS related to the phoneme x in the y-th section is given from the spectral envelope / inclination memory 11 by the spectral inclination conversion unit 8. The average spectral inclination of the recognition target speaker T given sT is read out, and the difference between the two average spectral inclinations is calculated. Then, as shown in FIG. 9B, the normalized spectral envelope of the recognition target speaker T is corrected by correcting the slope of the modified spectral envelope obtained in step S44 by the difference between the average spectral slopes. Desired. Thus, the spectrum envelope T of the recognition target speaker T is normalized to the spectrum envelope S of the standard speaker S. After that, the process proceeds to step S48.
[0077]
  Here, the frequency warping function memory 13 stores a plurality of average frequency warping functions for each phoneme, for each similar phoneme, for each voiced / unvoiced segment, and for each entire speech segment. Therefore, an appropriate average frequency warping function can be selected according to the amount of utterance data of the recognition target speaker T in the learning mode as follows. That is, when there is little or no utterance data of a certain phoneme (eg, back tongue vowel / o /), the average frequency warping of a similar phoneme (eg, back tongue vowel / a /) of the phoneme (/ o /) Select a function, or an average frequency warping function and average spectral slope of a voiced segment. Alternatively, when the utterance data of the phoneme (/ o /) is sufficiently large, the average frequency warping function and the average spectrum inclination of the phoneme (/ o /) are selected as described above. By doing so, it is possible to cope with a case where the amount of learning utterance data of the recognition target speaker T is small, and the utterance burden of the recognition target speaker T can be reduced.
[0078]
  In step S46, the learning data related to the phoneme x of the speaker T to be recognized is insufficient, and therefore the standard envelope S related to the phoneme x in the y-th section is retrieved from the frequency warping function memory 13 by the spectrum envelope converter 7. And the average frequency warping function of the similar phoneme, voiced / unvoiced section, or the entire speech section. Then, the spectrum envelope T of the recognition target speaker T loaded in step S42 is modified in the same manner as in step S44, and the modified spectrum envelope S is obtained.
[0079]
  In step S47, since the learning data related to the phoneme x of the recognition target speaker T is insufficient, the standard speaker related to the phoneme x in the y-th section is retrieved from the spectrum envelope / gradient memory 11 by the spectral gradient converter 8. The average spectral slope of the similar phonemes, voiced / unvoiced sections or the entire speech section of S and the recognition target speaker T is read. In the same manner as in step S45, the normalized spectrum envelope of the recognition target speaker T is obtained by correcting the slope of the modified spectrum envelope.
[0080]
  In step S48, based on the normalized spectrum envelope of the recognition target speaker T obtained in step S45 or step S47 by the speech recognition unit 3 and the time series of prosodic information such as the power, the HMM Voice recognition is performed again by the above. Then, the phoneme sequence of the obtained recognition result is output as a formal recognition result. In step S49, the phoneme section number y is the maximum value y._MAXIt is determined whether or not this is the case. As a result, the maximum value y_MAXIf smaller, the process proceeds to step S50, and if not, the processing operation in the recognition mode is terminated. In step S50, the phoneme section number y is incremented. After that, the process returns to step S42 to shift to the conversion of the next phoneme section into a modified spectrum envelope, the conversion into the normalized spectrum envelope, and the speech recognition process. In step S49, the phoneme section number y is the maximum value y._MAXIf it is determined as above, the processing operation in the recognition mode is terminated.
[0081]
  As described above, in the recognition mode, the spectral envelope and the spectral inclination are extracted for each phoneme section from the speech of the recognition target speaker T, and the phoneme number x and the speaker number sT are obtained by performing speech recognition and speaker recognition. obtain. Then, using the average frequency warping function between the recognition target speaker T and the standard speaker S corresponding to the phoneme number x and the speaker number sT obtained in the learning mode, the spectrum envelope of the recognition target speaker T is calculated. A transformed spectral envelope is obtained by converting into the spectral envelope of the standard speaker S. Furthermore, the modified spectral envelope is equal to the difference between the average spectral inclination of the recognition target speaker T corresponding to the phoneme number x and the speaker number sT obtained in the learning mode and the average spectral inclination of the standard speaker S. Correct the slope to obtain a normalized spectral envelope. Thereafter, the speech recognition unit 3 performs speech recognition again using the normalized spectrum envelope.
[0082]
  Thus, the input speech can be correctly recognized by using the spectrum envelope normalized to the standard speaker S. Further, the average frequency warping function used at that time is obtained by using the average spectrum envelope regarding the same speaker and the same phoneme as described above. Therefore, not only individual differences (hard differences) caused by physical differences but also individual differences (soft differences) caused by vocalizations can be normalized. However, phonological differences are not normalized.
[0083]
  In addition, the average frequency warping function used at the time of the spectral envelope conversion and the average spectral slope used at the time of correcting the slope of the modified spectral envelope are “phonemes”, “similar phonemes”, “voiced / unvoiced intervals”, ”Is required for each. Therefore, according to the amount of data of the recognition target speaker T stored in the spectrum envelope / inclination memory 11, the learning of the recognition target speaker T is performed by using the average frequency warping function and the average spectrum inclination in an appropriate category. Even when the amount of utterance data is small, speech recognition can be performed with high accuracy. That is, according to the present embodiment, it is possible to reduce the utterance burden of the learning target speaker.
[0084]
  In the above embodiment, the average spectrum envelope of the same phoneme of the speaker to be learned and the standard speaker is obtained in advance, and the average frequency warping function is obtained by performing nonlinear frequency axis spectrum matching using the average spectrum envelope. Looking for. However, the frequency warping function may be obtained by performing nonlinear frequency axis spectrum matching using individual spectral envelopes of the same phoneme, and the average frequency warping function may be obtained by averaging the frequency warping function within the same phoneme.
[0085]
  In the above embodiment, the spectral envelope is normalized only in the recognition mode. However, the present invention is not limited to this, and in the case of the speech recognition unit 3 that performs speech recognition by an unspecified speaker speech recognition method such as an HMM, it can be used in the learning mode as follows. I do not care. That is, after the frequency warping function is constructed as described above, the spectrum envelope is processed by the waveform analysis unit 1, the spectrum envelope / gradient extraction unit 2, the speaker recognition unit 5, and the spectrum envelope conversion unit 7 on the learning speech data. Normalization to the standard speaker direction is performed. Then, the phoneme standard pattern extracted from the obtained normalized spectrum envelope is stored in the phoneme standard pattern memory 4. However, in this case, the phoneme identification performance of the speech recognition unit 3 is not yet high enough, so that the phoneme labeling to the time series of the spectrum envelope is performed with supervision regardless of the recognition result of the speech recognition unit 3. Is preferable.
[0086]
  Thus, the phoneme identification performance of the speech recognition unit 3 can be improved by using the learned phoneme standard pattern extracted from the normalized spectrum envelope. Furthermore, when the speech recognition unit 3 uses the mixed distribution HMM, the variance of the phoneme standard pattern for each phoneme is reduced, and the number of distributions used for the mixed distribution of transition probabilities is suppressed to “1” or “less number”. be able to. Therefore, the memory capacity and processor load during execution can be reduced.
[0087]
  <Second Embodiment>
  FIG. 10 is a block diagram of the speech recognition apparatus according to the present embodiment. The vocal tract cross-sectional area extraction unit 22 extracts the vocal tract cross-sectional area from the glottis to the lips based on the result of autocorrelation analysis or covariance analysis in the waveform analysis unit 21. Based on the time series of the vocal tract cross-sectional area extracted by the vocal tract cross-sectional area extraction unit 22 and the prosodic information (power, pitch frequency, etc.) extracted by the waveform analysis unit 21, the speech recognition unit 23 Speech recognition is performed with reference to 24 phoneme standard patterns.
[0088]
  In the learning mode, the speaker recognizing unit 25 accumulates the input vocal tract cross-sectional area of the learning target speaker T and the time series of prosodic information in an internal memory or the like as a speaker voice standard pattern. On the other hand, in the recognition mode, speaker recognition is performed by comparing the time series of the vocal tract cross-sectional area and prosodic information of the input recognition target speaker T with the above-mentioned speaker voice standard pattern group.
[0089]
  The matching / average processing unit 26 calculates an average value for each phoneme, similar phoneme, voiced / unvoiced segment, and entire speech segment (speaker) with respect to the vocal tract cross-sectional area of each phoneme for each speaker. Do. In addition, the non-linear vocal tract axis matching by dynamic programming is used to match the average vocal tract cross-sectional area of the learning target speaker with the average vocal tract cross-sectional area of the standard speaker for each phoneme. Ask for. Further, an average value is calculated for the average vocal tract axis warping function of each phoneme for each of the similar phonemes, voiced / unvoiced sections and the entire speech section.
[0090]
  The vocal tract cross-sectional area conversion unit 27 uses the average vocal tract axis warping function between the recognition target speaker and the standard speaker to calculate the vocal tract cross-sectional area of the recognition target phoneme by the recognition target speaker. Convert to road cross section (normalized vocal tract cross section). Then, the voice recognition unit 23 performs voice recognition again using the normalized vocal tract cross-sectional area at the time of recognition. In this way, the speech recognition result normalized by the speaker is obtained.
[0091]
  Thus, in the second embodiment, instead of the spectral envelope in the first embodiment, the vocal tract cross-sectional area highly relevant to the spectral envelope is used, and the vocal tract cross-sectional area of the recognition target speaker is used. The vocal tract axis is nonlinearly expanded and contracted to obtain the vocal tract cross-sectional area of the standard speaker. Therefore, as in the case of the first embodiment, not only individual differences (hard differences) caused by physical differences but also individual differences (soft differences) caused by vocalizations are normalized with a small amount of audio data. In addition, speech recognition can be performed without normalizing the phoneme difference.
[0092]
  <Third Embodiment>
  FIG. 11 shows a block diagram of a speech recognition system using the speech recognition method in the first embodiment or the second embodiment. The speech recognition means 32 with the speaker normalization function corresponds to the speech recognition device in the first embodiment or the second embodiment. However, the obtained normalized spectrum envelope or normalized vocal tract cross-sectional area is not sent to the speech recognition unit. Therefore, the speech recognition unit does not perform a final speech recognition result using the normalized spectrum envelope or the normalized vocal tract cross-sectional area.
[0093]
  The transmission control unit 33 controls the connection between the speech recognition means 32 with the speaker normalizing function and the communication network 34. Here, the speech recognition means 32 with the speaker normalizing function and the transmission control unit 33 may be configured integrally, or may be separated if information can be transmitted by wire or wirelessly. Since the user connects the speech recognition means 32 with the speaker normalization function and the transmission control unit 33 as a set to the communication network 34, the voice recognition means 32 with the speaker normalization function and the transmission control unit 33 are grouped together. The user terminal 31 is called.
[0094]
  The communication network 34 may be either a LAN (Local Area Network) or a WAN (Wide Area Network). Further, any network of wired communication and wireless communication may be used. The center system 35 is accessed from the user terminal 31 via the communication network 34, performs various information processing, and returns a service to the user terminal 31.
[0095]
  The control unit 36 of the center system 35 controls the connection between the communication network 34 and the center system 35. The phoneme recognition unit 37 recognizes phonemes based on the speaker normalized parameters transmitted from the speech recognition means 32 with speaker normalization function via the communication network 34. The language processing unit 38 performs language processing on the phoneme sequence recognized by the phoneme recognition unit 37 using the vocabulary dictionary and grammar stored in the memory 39, and the speech processing unit 32 with speech normalization function has a speech. Recognizes the input language. The application 40 is a program that is activated under the control of the control unit 36 based on the recognition result of the language processing unit 38 and provides a service to the user terminal 31.
[0096]
  FIG. 12 shows a flowchart of the speaker normalized speech recognition / information processing operation performed by the user terminal 31 and the center system 35 in the speech recognition system having the above configuration. In the following description, the speech recognition means 32 with speaker normalization function functions in the same manner as in the first embodiment, and outputs a normalized spectrum envelope as the normalized parameter. Do
  Hereinafter, the speaker normalized speech recognition / information processing operation will be described in detail with reference to FIG. At that time, the speech recognition means 32 with the speaker normalization function of the user terminal 31 learns in the same manner as in the first embodiment, and the frequency warping function memory stores “phonemes” and “similar phonemes”. , It is assumed that the average frequency warping function for each of “voiced sound section / unvoiced sound section” and “speech section” is stored. Steps S51 to S54 are processing operations on the user terminal 31 side, and steps S55 to S59 are processing operations on the center system 35 side.
[0097]
  In step S51, the speech produced by the speaker T is input to the speech recognition means 32 with the speaker normalizing function. In step S52, the speech recognition means 32 with speaker normalization function analyzes the waveform of the input speech in the same manner as in the first embodiment, and obtains a time series of acoustic parameters such as cepstrum and power. It is done. In step S53, the average frequency warping function between the recognition target speaker T and the standard speaker S is further used based on the obtained time series of parameters as in the first embodiment. Speaker normalization is performed on the average spectral envelope, and a normalized spectral envelope is obtained. In step S54, the normalized spectrum envelope normalized by the speaker is transmitted to the control unit 36 of the center system 35 via the communication network 34 by the transmission control unit 33. In this embodiment, the spectral envelope is transmitted as an acoustic parameter. However, the spectral envelope is subjected to an inverse FFT (Fast Fourier Transform) process and extracted by performing another cepstrum or other process. An acoustic parameter may be transmitted.
[0098]
  In step S 55, the normalized spectral envelope time series transmitted via the communication network 34 is received by the control unit 36 of the center system 35 and sent to the phoneme recognition unit 37. In step S56, the phoneme recognition unit 37 performs phoneme recognition. Then, a phoneme lattice is obtained as a recognition result and returned to the control unit 36.
[0099]
  In step S57, the control unit 36 sends the received phoneme lattice to the language processing unit 38. The language processing unit 38 refers to the vocabulary and grammar related to the language lattice stored in the memory 39 and recognizes the word uttered by the speaker T. In step S58, a sentence uttered by the speaker T is further recognized based on the recognized word. Then, the recognized sentence is returned to the control unit 36. In step S59, the control unit 36 activates the application 40 according to the content of the received sentence. Then, for example, additional information is input or information is generated by the application 40, and the result is transmitted to the user terminal 31 by the control unit 36 via the communication network 34 to provide a service. After that, the speaker normalized speech recognition / information processing operation is terminated.
[0100]
  Thus, in the present embodiment, the normalized spectrum envelope or the normalized vocal tract cross-sectional area is provided on the user terminal 31 side in the same manner as the speech recognition apparatus in the first embodiment or the second embodiment. The voice recognition means 32 with a speaker normalizing function that can be obtained is installed. Then, the obtained normalized spectral envelope or normalized vocal tract cross-sectional area is transmitted to the center system 35 by the transmission control unit 33 via the communication network 34.
[0101]
  On the other hand, the phoneme recognition unit 36 and the language processing unit 38 are provided on the center system 35 side, and the user's voice to the user terminal 31 is based on the normalized spectrum envelope or the normalized vocal tract cross-sectional area. The input sentence is recognized, and the application 40 is activated according to the recognition result.
[0102]
  Therefore, compared to the case where the user terminal 31 recognizes the user's voice input sentence and transmits an instruction to start the application 40 according to the recognition result to the center system side via the communication network 34, There is no need to mount a language processing unit on the user terminal 31 side, and the configuration on the user terminal 31 side can be simplified.
[0103]
  In each of the above embodiments, a phoneme is used as the speech unit. However, the present invention is applicable to a syllable.
[0104]
  By the way, the waveform analysis units 1 and 21, the spectrum envelope / gradient extraction unit 2, the vocal tract cross-sectional area extraction unit 22, the speech recognition units 3 and 23, the speaker recognition units 5 and 25, and the nonlinear frequency axis spectrum in the above embodiments. The above speaker normalized speech recognition processing function by the matching unit 12, the averaging processing unit 14, the spectrum envelope conversion unit 7, the spectrum inclination conversion unit 8, and the vocal tract cross-sectional area conversion unit 27 is the speaker recorded in the program recording medium. This is realized by a normalized speech recognition processing program. Here, the program recording medium is a program medium comprising a ROM (Read Only Memory) or the like. Alternatively, it may be a program medium that is loaded into an external auxiliary storage device and read out. In any case, the program reading means for reading the speaker normalized speech recognition processing program from the program medium may be configured to directly access and read the program medium, or provided in the RAM. The program storage area (not shown) may be downloaded, and the program storage area may be accessed and read. It is assumed that a download program for downloading from the program medium to the program storage area of the RAM is stored in the main unit in advance.
[0105]
  Here, the program medium is configured to be separable from the main body side, and is a tape system such as a magnetic tape or a cassette tape, a magnetic disk such as a floppy disk or a hard disk, or a CD (compact disk) -ROM, MO (magneto-optical). Optical discs such as discs, MDs (mini discs), DVDs (digital video discs), card systems such as IC (integrated circuit) cards and optical cards, mask ROMs, EPROMs (ultraviolet erasable ROMs), EEPROMs (electrical This is a medium that carries a fixed program including a semiconductor memory system such as an erasable ROM) and a flash ROM.
[0106]
  Further, the program medium may be a medium that carries the program in a fluid manner by downloading from a communication network or the like. In this case, it is assumed that a download program for downloading from the communication network is stored in the main device in advance. Or it shall be installed from another recording medium.
[0107]
  It should be noted that what is recorded on the recording medium is not limited to a program, and data can also be recorded.
[0108]
【The invention's effect】
  As is clear from the above, the speech recognition apparatus of the first invention extracts the first and second spectral envelopes based on the utterances of the first and second speakers by the spectral envelope extracting means, and nonlinear frequency axis spectral matching means. To obtain a frequency warping function representing the correspondence between the frequency axes of the first and second spectrum envelopes, and the spectrum envelope conversion means converts the first spectrum envelope to be recognized into a spectrum related to the second speaker based on the frequency warping function. Since the input speech is recognized on the basis of the converted normalized spectral envelope by the speech recognition means, the frequency axis of the first spectral envelope to be recognized by the first speaker is set to the frequency axis of the frequency axis. Non-linear expansion and contraction can be performed according to the correspondence and normalized to the spectrum envelope by the second speaker.
[0109]
  Therefore, according to the present invention, when the first speaker is the unspecified speaker and the second speaker is the standard speaker, the spectrum envelope based on the utterance by the unspecified speaker is changed to the spectrum envelope of the standard speaker. Speaker normalization can be performed to perform highly accurate speech recognition. Furthermore, by obtaining the frequency warping function based on the average value of the plurality of first and second spectrum envelopes for the same voice unit, not only individual differences (hard differences) due to physical differences but also vocalization Individual differences (soft differences) can be normalized. In such a case, the phoneme difference is not normalized because it is nonlinearly expanded and contracted between the same phonemes.
[0110]
  Further, the frequency warping function is converted into phonemes by averaging means. , Similar phonemes , Voiced interval / An unvoiced sound segment and a speaker are grouped, an average value is calculated for each group, an average frequency warping function is obtained, and the spectrum envelope conversion means is used for the first speaker relating to the same phoneme as the phoneme of the recognition target first spectrum envelope Since the average frequency warping function of the group determined according to the amount of data is used as the frequency warping function, according to the amount of utterance data of the first speaker stored in the first memory means, An appropriate group of average frequency warping functions can be selected and used in place of the frequency warping function. Therefore, the utterance data of the first speaker ( Learning data ) High accuracy voice recognition even when the amount of Can reduce the burden on the speaker. Further, by obtaining the average frequency warping function for each phoneme, it is possible to normalize individual differences caused by vocal utterances.
[0111]
  The speech recognition apparatus according to the first aspect of the present invention extracts the first and second spectral envelope slopes based on the utterances of the first and second speakers by the spectral slope extraction means, and the first slope of the spectral slope correction means. Therefore, if the slope of the normalized spectrum envelope obtained by the spectrum envelope conversion means is corrected based on the difference in slope of the second spectrum envelope, the speaker normalization can be performed more accurately to the voice of the second speaker.The
[0112]
  MaThe speech recognition apparatus of the first invention uses the speech unit as a phoneme, and the averaging means determines the slope of the first spectrum envelope and the slope of the second spectrum envelope as phonemes, similar phonemes, voiced sound sections / unvoiced sound sections. And grouping for each speaker, calculating an average value for each group to obtain an average spectral inclination, and determining the average spectral inclination of any of the groups to which the phonemes of the recognition target first spectral envelope belong. If used as the slope of the spectral envelope, the average spectral slope of an appropriate group is selected according to the amount of utterance data of the first speaker stored in the first memory means, and the slope of the spectral envelope is selected. Can be used instead of Therefore, even when the amount of utterance data (learning data) of the first speaker is small, highly accurate speech recognition can be performed. Furthermore, by obtaining the average spectral inclination for each phoneme, it is possible to normalize individual differences caused by vocal utterances.
[0113]
  The speech recognition apparatus according to the first invention isRecordingThe spectrum inclination correcting means is used according to the data amount of the first speaker related to the recognition target phoneme stored in the first memory.RuheiIf a group of average spectral inclinations is determined, it is averaged within an optimal range according to the amount of learning data by the first speaker regarding the recognition target phoneme.TabiraAn average spectral slope can be used. Therefore, even when the amount of learning data of the first speaker is small, a high recognition rate can be obtained.
[0114]
  In the speech recognition apparatus according to the first aspect of the invention, when the speech recognition means obtains the frequency warping function, the extracted time series of the first spectrum envelope or the second spectrum envelope is recognized by the unspecified speaker speech. If the speech unit name is recognized by the method and the recognition result speech unit name is sent to the first memory means, the label speech unit name can be automatically obtained from the first and second spectrum envelopes. Therefore, it is possible to easily perform the labeling process for the spectrum envelope or the slope of the spectrum envelope.
[0115]
  In the speech recognition apparatus according to the first aspect of the invention, in the learning mode for obtaining the frequency warping function, the prosody information is extracted by analyzing the waveforms of a plurality of known speakers by the waveform analysis means, and the speaker recognition means While registering the prosodic information and the spectral envelope as a speech standard pattern, in the recognition mode, the speaker recognition means recognizes the speaker to be recognized using the registered speech standard pattern, and the recognized speaker If the frequency warping function or the first spectral inclination related to the name is used for normalization of the recognition target first spectral envelope, there are a plurality of speakers as the first speaker, and each first speaker is If the above-mentioned frequency warping function, first spectrum envelope and first spectrum slope are stored, speaker normalized speech recognition can be performed for unspecified speakers. You can.
[0116]
  In the speech recognition apparatus according to the first aspect of the invention, the speech recognition unit may perform speech recognition using a standard pattern extracted from the normalized spectrum envelope obtained by the spectrum envelope conversion unit. Thus, speech recognition can be performed with high identification accuracy using a standard pattern normalized by a speaker. Further, when the voice recognition means is a voice recognition means using a mixed distribution HMM, the variance of the phoneme standard pattern for each voice unit is reduced, and the number of distributions used for the mixed distribution of transition probabilities is set to “1” or “ It can be reduced to “less number”. Therefore, the memory capacity and processor load during execution can be reduced.
[0117]
  Further, the second invention extracts the first and second vocal tract cross-sectional areas based on the utterances of the first and second speakers by the vocal tract cross-sectional area extracting means, and the non-linear vocal tract axis matching means extracts the first and second vocal tract cross-sectional areas. A vocal tract warping function representing the correspondence between the vocal tract axes of the second vocal tract cross-sectional area is obtained, and the first vocal tract cross-sectional area to be recognized is determined based on the vocal tract axial warping function by the vocal tract cross-sectional area conversion means. Since the vocal tract cross-sectional area for the second speaker is converted to the vocal tract cross-sectional area, the vocal tract axis of the first vocal tract cross-sectional area to be recognized by the first speaker is nonlinearly expanded and contracted according to the correspondence of the vocal tract axis, Can be normalized to the vocal tract cross-sectional area of the person.
[0118]
  Therefore, according to the present invention, if the first speaker is the unspecified speaker and the second speaker is the standard speaker, the vocal tract cross-sectional area based on the utterance by the unspecified speaker is determined as the standard speaker's voice. It is possible to perform speech recognition with high accuracy by normalizing the speaker to the road cross-sectional area. Furthermore, by obtaining the above vocal tract axis warping function based on the average value of a plurality of first and second vocal tract cross-sectional areas for the same voice unit, not only individual differences due to physical differences but also vocalization Individual differences can also be normalized. In that case, only the vocal tract axis is nonlinearly expanded and contracted and the intensity axis is not expanded and contracted, so that the phoneme difference is not normalized. Therefore, it is possible to perform speech recognition with high individual difference normalization performance.
[0119]
  Further, the vocal tract axis warping function is converted into phonemes by averaging means. , Similar phonemes , Voiced interval / An unvoiced speech segment and a speaker are grouped, an average value is calculated for each group to obtain an average vocal tract axis warping function, and the vocal tract cross-sectional area conversion means performs the same phoneme as the phoneme of the recognition target first vocal tract cross-sectional area. Since the average vocal tract warping function of the group determined according to the amount of data of the first speaker related to the first speaker is used as the vocal tract warping function, the first stored in the first memory means is used. Depending on the amount of utterance data of the speaker, an appropriate group of average vocal tract warping functions can be selected and used instead of the vocal tract warping functions. Therefore, the utterance data of the first speaker ( Learning data ) Even when the amount of speech is small, highly accurate speech recognition can be performed, and the speaker's speech burden can be reduced. Furthermore, by obtaining the average vocal tract axis warping function for each phoneme, individual differences due to vocal utterances can be normalized.
[0120]
  According to a third aspect of the present invention, there is provided a speech recognition method for extracting first and second spectrum envelopes from the uttered sounds of the first and second speakers, and for extracting the first and second spectrum envelopes. Nonlinear frequency expansion / contraction matching using dynamic programming is performed to obtain the frequency warping function, a first spectrum envelope to be recognized is extracted, and the first spectrum envelope to be recognized is calculated based on the frequency warping function. Since the input speech is recognized on the basis of the converted spectral envelope related to the speaker and the normalized speech envelope is converted, only individual differences due to physical differences can be obtained as in the case of the first invention. It is also possible to normalize individual differences caused by vocalization. In this case, the phoneme difference is not normalized, and speech recognition with high individual difference normalization performance can be performed.
[0121]
  Further, the frequency warping function is converted into phonemes by averaging means. , Similar phonemes , Voiced interval / The first speaker with respect to the same phoneme as the phoneme of the first spectrum envelope to be recognized at the time of conversion of the spectrum envelope, by calculating an average value for each group and calculating an average frequency warping function. Since the average frequency warping function of the group determined according to the amount of data of the first speaker is used as the frequency warping function, the average of the appropriate group according to the stored amount of utterance data of the first speaker is used. A frequency warping function can be selected and used in place of the frequency warping function. Therefore, the utterance data of the first speaker ( Learning data ) Even when the amount of speech is small, highly accurate speech recognition can be performed, and the speaker's speech burden can be reduced. Further, by obtaining the average frequency warping function for each phoneme, it is possible to normalize individual differences caused by vocal utterances.
[0122]
  In the speech recognition method of the third invention, the slopes of the first and second spectrum envelopes are extracted from the uttered sounds of the first and second speakers, and the difference between the slopes of the first and second spectrum envelopes is extracted. If the normalized spectral envelope after the conversion is corrected based on the above and the input speech is recognized, the normalized spectral envelope after the inclination correction is used, and the voice of the second speaker is more accurately detected. The speaker can be normalized.
[0123]
  The speech recognition system according to a fourth aspect of the present invention is the user terminal side, wherein the normalized spectrum envelope or the normalized voice related to the input speech obtained by the speech recognition device according to the first aspect or the second aspect is provided. The road cross-sectional area is transmitted to the center system via the communication network by the transmission control means, and on the center system side, based on the normalized spectrum envelope or the normalized vocal tract cross-sectional area received by the control unit, Since the input speech recognition unit recognizes the input speech to the speech recognition device, speech recognition with high individual difference normalization performance can be performed as in the case of the first invention or the second invention. .
[0124]
  Furthermore, it is not necessary to perform input speech recognition processing and language processing on the user terminal side. Therefore, the configuration of the user terminal can be simplified.
[0125]
  According to a fifth aspect of the present invention, there is provided a program recording medium comprising: a computer, the spectral envelope extracting means, the nonlinear frequency axis spectral matching means, the spectral envelope conversion means, the voice recognition means,Averaging means ,Since a speaker normalization speech recognition processing program that functions as a spectrum inclination extraction means, a spectrum inclination correction means, and a speaker recognition means is recorded, the first recognition target by the first speaker is recorded using the frequency warping function. The frequency axis of the spectral envelope can be nonlinearly expanded and contracted to be converted into a spectral envelope by the second speaker. Furthermore, the slope of the converted normalized spectrum envelope can be corrected based on the difference between the slopes of the first and second spectrum envelopes. Therefore, the input speech can be recognized with high accuracy based on the normalized spectrum envelope after the tilt correction.
[0126]
  That is, according to this invention, as in the case of the first invention described above, not only individual differences caused by physical differences but also individual differences caused by vocalizations can be normalized. High voice recognition can be performed.
[Brief description of the drawings]
FIG. 1 is a block diagram of a speech recognition apparatus according to the present invention.
FIG. 2 is a detailed block diagram of a matching / averaging processor in FIG. 1;
FIG. 3 is a diagram illustrating an example of a spectral tilt and a spectral envelope.
4 is a flowchart in a learning mode of speaker normalized speech recognition processing operation by the speech recognition apparatus shown in FIG. 1;
FIG. 5 is a flowchart in the learning mode following FIG. 4;
FIG. 6 is a flowchart at the time of recognition mode in speaker normalization voice recognition processing operation;
FIG. 7 is a flowchart in the recognition mode following FIG. 6;
FIG. 8 is a diagram showing a concept of nonlinear frequency axis spectrum matching by dynamic programming.
FIG. 9 is a conceptual diagram of spectrum envelope normalization.
FIG. 10 is a block diagram of a speech recognition apparatus different from FIG.
FIG. 11 is a block diagram of a voice recognition system according to the present invention.
12 is a flowchart of speaker normalized speech recognition / information processing operations performed by the user terminal and the center system in FIG.
[Explanation of symbols]
  1, 21 ... Waveform analysis unit,
  2 ... spectrum envelope / tilt extraction unit, 3, 23 ... voice recognition unit,
  4,24 ... Phoneme standard pattern memory, 5,25 ... Speaker recognition unit,
  6, 26 ... matching and averaging processing unit, 7 ... spectrum envelope conversion unit,
  8: Spectral slope conversion unit, 11: Spectral envelope / tilt memory,
12 ... Nonlinear frequency axis spectrum matching unit,
13 ... Frequency warping function memory, 14 ... Averaging processing unit,
22 ... Vocal tract cross-sectional area extraction unit, 27 ... Vocal tract cross-sectional area conversion unit,
31 ... User terminal,
32. Speech recognition means with speaker normalization function,
33 ... Transmission control unit, 34 ... Communication network,
35 ... Center system, 36 ... Control part,
37 ... Phoneme recognition unit, 38 ... Language processing unit,
39 ... Memory, 40 ... Application.

Claims (16)

A speech recognition device with a speaker normalizing function for converting a voice of a first speaker into a voice of a second speaker and normalizing the speaker,
A spectral envelope extracting means for extracting a first spectral envelope from the first voice uttered by the first speaker, and extracting a second spectral envelope from the second voice uttered by the second speaker;
First memory means for storing the extracted first spectrum envelope and second spectrum envelope with a label in units of speech;
For the same label, non-linear frequency expansion / contraction matching using dynamic programming is performed on the first spectrum envelope and the second spectrum envelope stored in the first memory, and the correspondence between the frequency axes of both spectrum envelopes is performed. A non-linear frequency axis spectrum matching means for obtaining a frequency warping function representing the attachment;
Second memory means for storing the frequency warping function with a label in units of speech;
The first spectrum envelope of the recognition target extracted by the spectrum envelope extraction means is input, and the frequency warping function having the same voice unit name as the voice unit of the recognition target first spectrum envelope is read from the second memory. A spectral envelope converting means for converting the recognition target first spectral envelope into a spectral envelope relating to the second speaker based on the issued frequency warping function;
Speech recognition means for recognizing input speech based on the converted normalized spectral envelope ;
The frequency warping functions stored in the second memory means are grouped into phonemes , similar phonemes , voiced / unvoiced intervals and speakers based on the labels, and an average value of the frequency warping functions belonging to each group is calculated. The average frequency warping function obtained is provided with an averaging means for giving a label of each group name and storing it in the second memory means,
The spectrum envelope transforming unit calculates an average frequency warping function of a group determined according to the amount of data of the first speaker related to the same phoneme as the phoneme of the first spectrum envelope to be recognized stored in the first memory. A speech recognition apparatus characterized by being used as the frequency warping function . The speech recognition apparatus according to claim 1,
The nonlinear frequency axis spectrum matching means, when performing the nonlinear frequency expansion / contraction matching, between adjacent channels when dividing each spectrum envelope into a plurality of channels in a frequency band with respect to the first spectrum envelope and the second spectrum envelope. A speech recognition apparatus using a difference between output values of. The speech recognition apparatus according to claim 1 or 2,
The first memory means stores the slopes of the first spectrum envelope and the second spectrum envelope with labels of voice units, and stores them.
The first spectrum envelope slope is extracted from the first voice uttered by the first speaker, while the second spectrum envelope slope is extracted from the second voice uttered by the second speaker and stored in the first memory means. Spectral tilt extraction means for storing;
The slope of the first spectrum envelope and the slope of the second spectrum envelope having the same speech unit name as the speech unit of the first spectrum envelope to be recognized are read from the first memory, and the spectrum envelope transformation is performed based on the difference between the slopes. Spectral tilt correction means for correcting the slope of the normalized spectrum envelope obtained by the means,
The speech recognition apparatus according to claim 1, wherein the speech recognition means recognizes an input speech based on a normalized spectrum envelope whose inclination is corrected. The speech recognition apparatus according to claim 3 ,
The speech unit is a phoneme,
The slopes of the first spectrum envelope and the second spectrum envelope stored in the first memory means are grouped into phonemes, similar phonemes, voiced / unvoiced intervals and speakers based on the labels. An average means for calculating the average value of the slope of the spectrum envelope to which the average spectrum slope belongs is assigned to each speaker name and each group name and stored in the first memory means.
The spectral inclination correcting means uses an average spectral inclination of any group to which the same phoneme as the recognition target first spectral envelope phoneme belongs as the spectral envelope inclination. . The speech recognition apparatus according to claim 4 ,
The spectrum inclination correction means determines a group of average spectrum inclination to be used according to the amount of data of the first speaker related to the same phoneme as the phoneme of the recognition target first spectrum envelope stored in the first memory. A speech recognition apparatus characterized by being configured to be configured as described above. The speech recognition apparatus according to any one of claims 1 to 3 ,
The speech recognition means recognizes the extracted time spectrum of the first spectrum envelope or the second spectrum envelope by the unspecified speaker speech recognition method when obtaining the frequency warping function, and the speech unit name of the recognition result is A speech recognition apparatus characterized by being sent to a first memory means . The speech recognition device according to claim 1 ,
The speech recognition apparatus according to claim 1, wherein the averaging means calculates the average frequency warping function by performing linear conversion between frequency warping functions to be averaged . The speech recognition apparatus according to claim 7 .
The frequency warping function is a lattice point corresponding to a DP path on a plane composed of the first and second spectrum channels when the first spectrum envelope and the second spectrum envelope are divided into a plurality of channels in the same frequency band. A matrix-like data format in which different element values are given to other grid points,
The linear transformation between the frequency warping functions is obtained by calculating a sum of element values of the same lattice points in a plurality of matrices corresponding to the frequency warping function to be the average value calculation, and using the obtained values as element values. speech recognition apparatus according to claim to Rukoto and the average frequency warping function. The speech recognition apparatus according to claim 8.
In the case where the spectrum envelope converting means converts the intensity into a frequency band having the second spectrum envelope, at the grid point of the row or column relating to the corresponding channel of the second spectrum envelope in the matrix of the average frequency warping function to be used. The product sum of the element value of each lattice point and the intensity in the channel of the first spectrum envelope corresponding to the lattice point is obtained, and the value of this product sum is set as the intensity in the frequency band of the second spectrum envelope. A voice recognition device characterized by the above. The speech recognition apparatus according to any one of claims 1 to 3 ,
There are multiple first speakers,
Waveform analysis means for extracting the prosodic information by performing waveform analysis on the voices of the first speaker and the second speaker;
In the learning mode for obtaining the frequency warping function, the prosodic information and spectrum envelope extracted from the speech of a plurality of known speakers are registered as speech standard patterns with each speaker's label, while unspecified speakers In the recognition mode for recognizing the speech of the recognition target speaker, the recognition target speaker is recognized using the registered speech standard pattern based on the prosodic information and the spectrum envelope extracted from the speech of the recognition target speaker. A speaker recognition means for outputting the speaker name of
It said spectrum transformation unit and the spectral tilt compensation means, the recognition mode, speech recognition, characterized that you have made to read the frequency warping function or the first spectral tilt related talker name from the speaker recognition unit apparatus. The speech recognition apparatus according to any one of claims 1 to 3,
The speech recognition means performs speech recognition by an unspecified speaker speech recognition method using a standard pattern extracted from the normalized spectrum envelope obtained by the spectrum envelope conversion means. Voice recognition device. A speech recognition device that normalizes a speaker by converting speech with the voice quality of the first speaker into speech with the voice quality of the second speaker ,
Vocal tract cross-sectional area extracting means for extracting a first vocal tract cross-sectional area from the first voice uttered by the first speaker and extracting a second vocal tract cross-sectional area from the second voice uttered by the second speaker; ,
First memory means for storing the extracted first vocal tract cross-sectional area and the second vocal tract cross-sectional area with a label of a voice unit;
For the same label, non-linear vocal tract expansion / contraction matching using dynamic programming is performed on the first vocal tract cross-sectional area and the second vocal tract cross-sectional area stored in the first memory, and both voices are obtained. A non-linear vocal tract axis matching means for obtaining a vocal tract axis warping function representing a correspondence between vocal tract axes of the tract cross-sectional area;
Second memory means for storing the vocal tract axis warping function with a label in units of speech;
The recognition target first vocal tract cross-sectional area extracted by the vocal tract cross-sectional area extraction means is input, and the vocal tract axis warping function having the same speech unit name as the speech unit of the recognition target first vocal tract cross-sectional area is described above. A vocal tract cross-sectional area conversion means for reading from the second memory and converting the recognition target first vocal tract cross-sectional area into a vocal tract cross-sectional area related to the second speaker based on the read vocal tract axis warping function;
Speech recognition means for recognizing input speech based on the converted normalized vocal tract cross-sectional area;
The vocal tract axis warping functions stored in the second memory means are grouped into phonemes , similar phonemes , voiced / unvoiced sections and speakers based on the labels, and the average of the vocal tract axis warping functions belonging to each group. An averaging means for calculating a value and adding the obtained average vocal tract axis warping function with a label of each group name and storing it in the second memory means,
The vocal tract cross-sectional area conversion means is an average of groups determined according to the amount of data of the first speaker related to the same phoneme as the phoneme of the recognition target first vocal tract cross-sectional area stored in the first memory. A speech recognition apparatus , wherein a vocal tract axis warping function is used as the vocal tract axis warping function . A speech recognition method with a speaker normalization function for speaker normalization the voice is converted into voice of the second speaker of the first speaker,
Extracting a first spectral envelope from the first speech uttered by the first speaker, while extracting a second spectral envelope from the second speech uttered by the second speaker ;
Respect same speech unit name, performing a nonlinear frequency expansion matching using dynamic programming with respect to the above extracted the first spectral envelope and the second spectral envelope, corresponding frequency axis in both spectral envelope determining a frequency warping function representing Paste,
Extracting a first spectral envelope of a recognition target;
And converting on the basis of the frequency warping function with the same speech unit name and the extracted recognition target first spectral envelope of the speech unit, the recognition target first spectral envelope to the spectrum envelope about the second speaker,
Recognizing input speech based on the transformed normalized spectral envelope ;
The frequency warping function is grouped into phonemes , similar phonemes , voiced / unvoiced intervals and speakers based on the labels, and the average value of the frequency warping functions belonging to each group is calculated. Providing a label for each group name to
The recognition target first spectral envelope steps of converting the spectral envelope for the second speaker, depending on the amount of data of the first speaker of the extracted related to the same phoneme phoneme of said recognition target first spectral envelope The speech recognition method, wherein an average frequency warping function of the determined group is used as the frequency warping function . Te speech recognition method smell of claim 13,
While the first speaker to extract the inclination of the first spectral envelope from a first speech uttered, comprising the steps of the second speaker to extract the inclination of the second spectral envelope from the second voice uttered,
Based on the difference between the slope of the first spectrum envelope and the slope of the second spectrum envelope having the same voice unit name as the speech unit of the recognition target first spectrum envelope, the slope of the converted normalized spectrum envelope is corrected. With steps ,
In the step of recognizing the input speech, the normalized spectral envelope after the inclination correction is used . A voice recognition system including a user terminal capable of exchanging voice information via a communication network and a center system,
In the user terminal,
A speech recognition device according to any one of claims 1 to 12,
While comprising transmission control means for transmitting the normalized spectrum envelope or the normalized vocal tract cross-sectional area related to the input speech obtained by the speech recognition device to the center system via the communication network,
In the center system,
A control unit that receives the normalized spectrum envelope or the normalized vocal tract cross-sectional area transmitted from the user terminal via the communication network;
A speech recognition system comprising: an input speech recognition unit that recognizes an input speech to the speech recognition device based on the normalized spectrum envelope or the normalized vocal tract cross-sectional area . Computer
The spectrum envelope extraction means , nonlinear frequency axis spectrum matching means , spectrum envelope conversion means , speech recognition means , averaging means in claim 1, spectrum inclination extraction means , spectrum inclination correction means in claim 3, and story in claim 10 A computer-readable program recording medium on which a speaker normalized speech recognition processing program that functions as a person recognition means is recorded .

Images