JP4981850B2 - Voice recognition apparatus and method, program, and recording medium - Google Patents

Description

  The present invention relates to a speech recognition apparatus and method, a program, and a recording medium for efficiently recognizing speech data of various sound qualities.

  In recent years, it has become possible to easily obtain a large amount of audio data as a memory element for recording audio data becomes cheaper. When recognizing such audio data, there arises a problem that the recognition accuracy and processing time greatly vary depending on the quality of the audio data.

  FIG. 10 shows a functional configuration of a conventional speech recognition apparatus 900. The speech recognition apparatus 900 includes an A / D conversion unit 90, a feature amount analysis unit 91, a speech recognition processing unit 92, an acoustic model parameter memory 93, and a language model parameter memory 94.

The A / D conversion unit 90 converts the sound of the input analog signal into a discrete digital signal with a sampling frequency of 16 kHz, for example. The feature quantity analysis unit 91 receives the voice digital signals that have been converted into discrete values, and calculates a voice feature quantity O _t for each frame, for example, with 320 voice digital signals as one frame (20 ms). The voice feature amount O _t is calculated, for example, by Mel frequency cepstrum coefficient (MFCC) analysis.

Based on the beam search algorithm, the speech recognition processing unit 92 refers to the acoustic model recorded in the acoustic model parameter memory 93 and the language model recorded in the language model parameter memory 94 with the speech feature amount O _t as an input. Output the voice recognition result. The beam search algorithm is a predetermined number that can roughly guarantee the existence of a recognition result having the highest cumulative likelihood from the highest cumulative likelihood (sum of the likelihood of the acoustic model and the likelihood of the language model) for each frame. This is a procedure for searching while leaving a speech recognition result candidate (hypothesis) of (beam width).

  Here, conventionally, the acoustic model is adaptively changed to cope with a change in the quality of the audio data (Non-Patent Document 1). FIG. 11 shows the idea. Estimate background noise and audio distortion included in audio data recorded in the field, and generate a transformation matrix that changes the standard acoustic model on which adaptation is based. Then, the standard acoustic model is multiplied by the transformation matrix to adapt the acoustic model to the environment in the field. By adapting the acoustic model to various environments in this way, it has been possible to cope with fluctuations in the quality of audio data.

Masakazu Masahiro, 5 others, “Free Speech Recognition Technology that Listens to Natural Conversations with Customers“ VoiceRex ”, NTT Technical Journal, pp.15-18, 2006.11

  Conventionally, the idea of adapting the above-described acoustic model has been common as a method for dealing with the quality of variously changing audio data. In other words, this is the idea that the speech recognition device side should cope with a change in the quality of the speech data. As a result, for excessively distorted speech data and the like, even if the acoustic model is adapted, there is not a sufficient difference in the likelihood between hypotheses during the beam search, and the search efficiency becomes poor and the processing time is reduced. Increase. As a result, there is a problem that it takes a long time and a highly accurate recognition result cannot be obtained.

  The present invention has been made in view of such problems, and a speech recognition apparatus that evaluates the quality of speech data as a preliminary process of speech recognition processing and controls the operation of the recognition processing based on the evaluation result. Another object of the present invention is to provide a speech recognition apparatus that efficiently recognizes a plurality of sound files based on the idea, a method, a program thereof, and a recording medium.

The speech recognition apparatus according to the present invention includes a feature amount analysis unit, a frame sound quality estimation unit, an average sound quality estimation unit, a speech recognition processing control unit, and a speech recognition processing unit. The feature amount analysis unit analyzes the speech feature amount of the speech digital signal included in the input speech file in units of frames. The frame sound quality estimation unit refers to the GMM for each frame, calculates a GMM likelihood corresponding to the sound feature amount of the frame, and outputs the GMM likelihood as the frame sound quality . The average sound quality estimation unit calculates a sound quality level that is the sound quality of the sound file from the frame sound quality of all frames of the sound file. The voice recognition processing control unit outputs a control signal including a non-recognition instruction signal indicating that voice recognition processing is not performed when the sound quality level is lower than a predetermined threshold. The voice recognition processing unit does not perform voice recognition processing on the voice file when the control signal includes a non-recognition instruction signal.

  In addition to the functional configuration described above, the voice recognition device that efficiently recognizes a plurality of voice files further includes a voice file control unit, a voice file processing unit, and a voice file memory. The audio file control unit inputs the audio file information of the audio digital signal and the control signal and determines the processing order of the audio file information. The voice file processing unit records the voice digital signal in the voice file memory in units of the voice file, and outputs the voice digital signal recorded in the order of processing to the voice recognition processing unit.

  According to the speech recognition apparatus of the present invention, since the operation of the speech recognition processing unit is adaptively changed by the control signal corresponding to the quality of the speech data, it is possible to improve the processing time efficiency while maintaining the speech recognition accuracy. I can do it. Moreover, in the voice recognition apparatus of the present invention that performs voice recognition of a plurality of voice files, it is possible to process the voice files in order of high voice quality based on the control signal. In addition, since it is possible to exclude those whose voice quality does not reach a predetermined level from the recognition target, there is an effect of improving the processing efficiency of the whole voice recognition process. That is, since the voice data with poor quality does not become a bottleneck, the efficiency of the voice recognition process is improved.

The figure which shows the function structural example of the speech recognition apparatus 100 of this invention. The figure which shows the operation | movement flow of the speech recognition apparatus 100. The figure which shows the function structural example of the frame sound quality estimation part 10. FIG. The figure which shows the function structural example of the frame sound quality estimation part 11. FIG. The figure which shows the setting method of the beam search width | variety of the control signal of the speech recognition process control part. The figure which shows the function structural example of the speech recognition apparatus 200 of this invention. The figure which shows the operation | movement flow of the speech recognition apparatus 200. The figure which shows the function structural example of the speech recognition apparatus 300 of this invention. The figure which shows the operation | movement flow of the speech recognition apparatus 300. The figure which shows the function structure of the conventional speech recognition apparatus 900. The figure which shows the idea corresponding to the fluctuation | variation of the quality of the audio | speech data disclosed by the nonpatent literature 1. FIG.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a functional configuration example of the speech recognition apparatus 100 of the present invention. The operation flow is shown in FIG. The speech recognition apparatus 100 includes a feature amount analysis unit 91, a frame sound quality estimation unit 10, an average sound quality estimation unit 20, a speech recognition processing control unit 30, a speech recognition processing unit 92 ′, an acoustic model parameter memory 93, A language model parameter memory 94 and a control unit 35 are provided. The feature amount analysis unit 91, the acoustic model parameter memory 93, and the language model parameter memory 94 are the same as those of the conventional speech recognition apparatus 900. The speech recognition processing unit 92 ′ differs from the speech recognition apparatus 92 in other operations except that it performs speech recognition processing based on the control signal output from the speech recognition processing control unit 30. The point that the A / D conversion unit 90 is provided when analog signal voice data is input is the same as the voice recognition apparatus 900.

  The speech recognition apparatus 100 is realized by reading a predetermined program into a computer configured with, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  Compared with the conventional speech recognition apparatus 900, the speech recognition apparatus 100 is new in that it includes a frame sound quality estimation unit 10, an average sound quality estimation unit 20, and a speech recognition processing control unit 30. In the following description, the description will focus on the different parts.

The feature amount analysis unit 91 receives the speech digital signal converted into a discrete value, and calculates the speech feature amount O _t for each frame in which a predetermined number of speech digital signals are one frame (for example, 20 ms) (step S91). ). The frame sound quality estimation unit 10 evaluates the sound quality of the audio digital signal for each frame (t) and outputs the frame sound quality q (t) (step S10).

The average sound quality estimation unit 20 estimates the sound quality level Q _T from the frame sound quality q (t) of a plurality of frames (step S20). The subscript T is a serial number of a plurality of frames.

Speech recognition processing control unit 30 sets the control signal at the time of speech recognition on the basis of the quality level Q _T (step S30). A specific example of the control signal will be described later. The voice recognition processing unit 92 ′ performs voice recognition processing based on the control signal set by the voice recognition processing control unit 30 (step S92 ′).

  The above operation is repeated until completion for all frames (N in step S35). The control unit 35 controls the operation and repetitive operation of each unit of the speech recognition apparatus 100. The control unit 35 may not only perform processing for each frame but also control each unit so that the above-described operation is performed for each audio file or each utterance unit.

According to the speech recognition apparatus 100, the speech recognition processing unit 92 adaptively performs recognition processing according to the control signal set by the speech recognition processing control unit 30. In other words, by setting the control signal corresponding to the quality level Q _T of a plurality of frames, it is possible to improve the efficiency of the processing time while maintaining the accuracy of speech recognition. Next, the operation of the speech recognition apparatus 100 will be described in more detail by showing a specific configuration example of each unit.

[Frame sound quality estimation unit]
FIG. 3 shows a functional configuration example of the frame sound quality estimation unit 10. The frame sound quality estimation unit 10 includes, for example, a GMM likelihood calculation unit 101 and a GMM (Gaussian Mixture Model) 102. The GMM 102 may be stored in the acoustic model parameter memory 93. The GMM likelihood calculating means 101 receives the speech feature amount O _t and refers to the GMM 102 to calculate a GMM likelihood corresponding to the speech feature amount O _t representing the frame sound quality q (t). For example, since the GMM 102 is learned from all phonemes of the learning data of the acoustic model, the likelihood indicates the degree of coincidence between the acoustic model and the speech feature amount O _t , and each frame is represented by the GMM likelihood value q (t). It is possible to evaluate the sound quality (whether or not it matches the acoustic model). That is, if the GMM likelihood is large, it can be evaluated that the sound quality is good (speech recognition accuracy is high), and if the value is small, the sound quality is bad (speech recognition accuracy is low).

  Instead of the GMM likelihood, an output probability value before conversion to a logarithmic value may be used. Furthermore, silence may be removed from the learning data of the GMM 102 and the GMM 102 may be used as a voice GMM. Further, both the speech GMM and the pause (silence) model may be collated, and the likelihood value with the higher likelihood may be used.

FIG. 4 shows a frame sound quality estimation unit 11 of another functional configuration example. The frame sound quality estimation unit 11 includes a power calculation unit 111, a voice / non-speech section detection unit 112, and an S / N calculation unit 113. The power calculation unit 111 calculates the power of each frame from the audio feature amount O _t . The voice / non-voice section detecting means 112 detects, for example, a frame having a power of a certain value or more as a voice section. Further, the likelihood of the speech GMM and the pause model may be compared, and a section having a high likelihood of the speech GMM may be set as a speech section. The S / N calculation means 113 calculates an S / N ratio that is a ratio of the power of the speech section to the non-speech section. This S / N ratio becomes the frame sound quality q (t).

[Average sound quality estimation section]
The average sound quality estimation unit 20 estimates the sound quality level Q _T by averaging, for example, GMM likelihood and S / N ratio, which are the frame sound quality q (t) output from the frame sound quality estimation unit 10 over a plurality of frames ( Formula (1)).

  Here, t is a frame number and T is the number of a plurality of frames. The subscript T is a serial number of the plurality of frames.

[Voice recognition processing control unit]
Speech recognition processing control unit 30 outputs a control signal as an input sound level Q _T. As a specific example of the control signal, for example, a beam search width N (Q _T ) can be considered. An example is shown in equation (2).

FIG. 5 illustrates the relationship between the sound quality level Q _T and the beam search width N (Q _T ). The horizontal axis is the sound quality level Q _T , and the vertical axis is the beam search width N (Q _T ).

As shown in FIG. 5, the expression (2) represents the beam search width N (Q _T ) (N _{min to} N _max ) corresponding to the sound quality level Q _T (Q _{min to} Q _max ) in a predetermined range as the sound quality level Q. _The idea is to proportionally distribute by the value of _T. Here, since the proportionality coefficient is a negative value, the sound quality level Q _T is small, the beam search width N (Q _T ) is large, Q _T is large, and N (Q _T ) is small. Of course, the relationship between the sound quality level Q _T and the beam search width N (Q _T ) may be a relationship that can be expressed by a non-linear function. When the beam search width N (Q _T ) is used as the control signal, the beam search width is not limited to the number beam width, and for example, the score beam width, the word end score beam width, the word end number beam width, etc. It may be.

Here, regarding the sound quality ranges such as the S / N ratio and the GMM likelihood, for example, Q _max and Q _min may be set to the maximum / minimum values from the sound quality distribution for the acoustic model learning data. Further, when the sound quality is set to the S / N ratio, a predetermined range such as Q _max = 30 [dB] and Q _min = 10 [dB] may be used. Regarding the beam search width, for example, N _max may be 1.5 times the beam width that is normally used, N _min may be half the beam width that is normally used, and the like.

In addition, when the sound quality level is extremely bad (for example, Q _T <Q _min ), even if the beam search width is increased, it is not possible to improve the accuracy and it takes much processing time. Therefore, the beam search width is reduced to, for example, N _min . May be. Further, the speech recognition process may not be performed by including the non-recognition instruction signal in the control signal.

[Voice recognition processing unit]
The speech recognition processing unit 92 ′ receives the speech feature quantity O _t and the beam search width N (Q _T ) of the control signal as input, and the acoustic model recorded in the acoustic model parameter memory 93 and the language model parameter memory 94. The speech recognition result is output based on the beam search algorithm with reference to the language model. The speech recognition processing unit 92 ′ is different from the conventional speech recognition device only in searching for a correct answer from beam search width N (Q _T ) speech recognition result candidates. That is, the operation of the speech recognition processing unit is adaptively changed. Since the beam search method itself is the same as that of a conventional speech recognition apparatus, detailed description thereof is omitted.

As described above, the speech recognition apparatus 100 adaptively changes the speech recognition processing by the quality level Q _T. In the example shown in FIG. 5, when the quality level Q _T is poor beam search width N (Q _T) large and the small beam search width N (Q _T) when quality level is good. That is, when the sound quality is good, a likelihood difference is added between the speech recognition result candidates (hypotheses), so that the speech recognition accuracy does not deteriorate even if the beam search width is narrowed, and the processing speed can be improved. On the other hand, if the sound quality is poor, it is difficult to add a likelihood difference between the speech recognition result candidates (hypotheses), so it is possible to improve the speech recognition accuracy by widening the beam width. However, if the sound quality is extremely poor, there is no likelihood difference between the speech recognition result candidates (hypotheses) even if the beam search width is widened. Conversely, the beam search width is narrowed or excluded from the speech recognition target. Thus, the processing speed can be improved. Accordingly, it is possible to improve the processing time efficiency while maintaining the voice recognition accuracy.

When the idea of adaptively changing the control signal described in the first embodiment according to the sound quality level Q _T is applied to a voice recognition device that recognizes a plurality of voice files, the voice recognition of the plurality of voice files can be performed efficiently. Is possible.

  FIG. 6 shows a functional configuration example of the speech recognition apparatus 200. The operation flow is shown in FIG. The voice recognition device 200 is different from the voice recognition device 100 in that it further includes a voice file control unit 40, a voice file processing unit 50, and a voice file memory 60. Other functional configurations are the same as those of the speech recognition apparatus 100.

The sound file control unit 40 receives sound file information (for example, sound file name) of a sound digital signal input from the outside, a sound quality level Q _T of the sound digital signal, and a control signal N (Q _T ) as an sound file. Is determined (step S40, FIG. 7). The audio file processing unit 50 records the feature amount of the audio digital signal in the audio file memory 60 in units of frames for each audio file (step S501). At the same time, a control signal is recorded. And the feature-value and control signal of the audio | voice digital signal recorded in the processing order determined by the audio | voice file control part 40 are output (step S50).

  The process of step S91 for analyzing the feature value and the process of step S501 for recording the feature value and the control signal in the sound file memory 60 in units of sound files are performed until all the input sound files are completed (N in step S502). ). Then, the feature amount and the control signal are output to the speech recognition processing unit 92 ′ in units of frames in accordance with the processing order of each file (step S503).

  The operation of the speech recognition processing unit 92 ′ is the same as that of the speech recognition apparatus 100 except that the control signal is given from the audio file processing unit 50. The voice recognition processing unit 92 ′ performs voice recognition processing based on the control signal (step S92 ′). The voice recognition process is repeated until all input files are completed (N in step S36), and the control unit 36 controls the operation.

Audio file control unit 40 determines the processing order on the basis of the quality level Q _T. The order of processing, if the descending order of quality level Q _T, will be the speech recognition processing by the voice recognition processing section 92 'from the good files sound sequentially. As a result, the voice files are subjected to voice recognition processing in order of good sound quality, so that a plurality of voice files can be efficiently recognized.

Further, the number and specifications of the computer to perform a speech recognition process, the total in the case where the speech recognition processing can not be performed on the file, by referring to the quality level Q _T, the speech recognition target only good voice file quality It can be.

The voice file processing section 50 comprises a sound quality range determining unit 501, the quality level Q _T in the quality range determining means 501 determines whether or not greater than the predetermined value Q _th, the audio file if smaller than the predetermined value May be discarded.

Although the audio file processing unit 50 has been described with respect to the example in which the feature amount is recorded in the audio file memory 60, the audio digital signal before the analysis of the feature amount may be recorded in the audio file memory 60. Further, the sound quality range determination unit 501 may select the discard file based on the predetermined value _Qth , for example, with the minimum value of the sound quality level for the acoustic model learning data as the predetermined value _Qth .

Further, without limiting to the minimum value of the sound quality level of the learning data, a predetermined value is obtained from the average μ and standard deviation σ of the sound quality level Q _T distribution when the sound quality level distribution of the learning data follows a normal distribution. It may be set to 2σ. Further, the voice recognition processing unit 92 ′ may be a general voice recognition device. In that case, no control signal is required, and the speech recognition apparatus performs speech recognition processing in order of good sound quality.

  FIG. 8 shows a functional configuration example of the speech recognition apparatus 300 of the present invention. The operation flow is shown in FIG. The speech recognition apparatus 300 further includes an unsupervised adaptation unit 80, an after-adaptation acoustic model parameter memory 95, and a second speech recognition processing unit 96 in addition to the functional configuration of the speech recognition apparatus 100 of the first embodiment. The speech recognition process is performed using an acoustic model learned by using the speech recognition result obtained by performing the speech recognition process in 100 as an adaptive label.

  The unsupervised adaptation unit 80 learns the acoustic model recorded in the acoustic model parameter memory 93 using the speech recognition result output from the speech recognition processing unit 92 ′ as an adaptation label, and generates an adaptive acoustic model (step S80, FIG. 9). The adaptive acoustic model is recorded in the post-adaptation acoustic model parameter memory 95.

  The second speech recognition processing unit 96 refers to the post-adaptation acoustic model parameter memory 95 and the language model parameter memory 94, and outputs a speech recognition result based on the beam search algorithm (step S96). The process of the second speech recognition process in step S96 may be the process of the speech recognition apparatus 100 of the first embodiment or may be the process of a general speech recognition apparatus. Note that, as the control signal is input to the unsupervised adaptation unit 80 by a broken line, the unsupervised adaptation unit 80 uses the speech recognition result output from the speech recognition processing unit 92 'as the adaptive label according to the value of the control signal. Whether or not to do so may be determined.

  As described above, according to the speech recognition apparatus 300, the acoustic model is learned using the result of speech recognition according to the sound quality level of the speech data as an adaptive label, so that the accuracy of the acoustic model can be improved. Then, it is possible to perform speech recognition processing using the highly accurate acoustic model. Further, according to the speech recognition apparatuses 100 and 200 of the present invention, the operation of the speech recognition processing unit is changed by the control signal according to the quality of the speech data, so that the efficiency of the speech recognition processing can be improved.

  In addition, although the frame sound quality estimation unit 10 according to the first embodiment has been described as an example configured with a GMM, a monophone likelihood may be calculated for each frame, and the frame sound quality may be estimated using the monophone likelihood. That is, all the monophones belonging to the acoustic model may be checked against the input feature quantity, and the sound quality may be evaluated using the maximum likelihood monophone with the highest likelihood.

Further, the processes described in the above method and apparatus are not only executed in time series according to the order of description, but also may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

  Further, when the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (11)

A feature quantity analysis unit that analyzes voice feature quantities of a voice digital signal included in an input voice file in units of frames;
A frame sound quality estimation unit that evaluates the sound quality of the audio digital signal for each frame and outputs a frame sound quality that is the sound quality of the frame;
An average sound quality estimation unit that calculates a sound quality level that is the sound quality of the sound file from the frame sound quality of all frames of the sound file;
A speech recognition processing control unit that outputs a control signal for controlling an operation during speech recognition processing based on the sound quality level;
A speech recognition processing unit that performs speech recognition processing based on the control signal;
Comprising
The frame sound quality estimation unit calculates a GMM likelihood corresponding to the voice feature amount of the frame with reference to the GMM, and outputs the GMM likelihood as the frame sound quality.
The voice recognition processing control unit outputs a control signal including a non-recognition instruction signal indicating that voice recognition processing is not performed when the sound quality level is lower than a predetermined threshold,
The voice recognition processing unit does not perform voice recognition processing of the voice file when the control signal includes a non-recognition instruction signal.
A speech recognition apparatus characterized by that. The speech recognition apparatus according to claim 1,
  The frame sound quality estimation unit uses a speech GMM obtained by removing silence from the GMM learning data, calculates a speech GMM likelihood that is a likelihood corresponding to a speech feature amount of the frame in the speech GMM, and the speech GMM Output the likelihood as the frame sound quality,
  Or
  The speech feature amount of the frame is compared with the silence model, and a silence likelihood that is a likelihood corresponding to the speech feature amount of the frame in the silence model is calculated, and is higher in the speech GMM likelihood and the silence likelihood. The likelihood value of the other is output as the frame sound quality,
  A speech recognition apparatus characterized by that. The speech recognition apparatus according to claim 1 or 2 ,
Furthermore, an audio file memory for recording a plurality of audio files,
An audio file control unit for determining the processing order of the audio file in descending order of the sound quality level, by inputting the audio file information of the audio digital signal, the sound quality level, and the control signal;
A voice file processing unit for recording the voice digital signal in the voice file memory in units of the voice file and outputting the recorded voice digital signal to the voice recognition processing unit in the processing order;
A speech recognition apparatus comprising: A feature quantity analysis unit that analyzes voice feature quantities of a voice digital signal included in an input voice file in units of frames;
A frame sound quality estimation unit that evaluates the sound quality of the audio digital signal for each frame and outputs a frame sound quality that is the sound quality of the frame;
An average sound quality estimation unit that calculates a sound quality level that is the sound quality of the sound file from the frame sound quality of all frames of the sound file;
A speech recognition processing control unit that outputs a control signal for controlling an operation during speech recognition processing based on the sound quality level;
A speech recognition processing unit that receives the speech feature value and the control signal as input and outputs a result of speech recognition processing based on the speech feature value as an adaptive label;
An unsupervised adaptation unit that learns an acoustic model using the adaptation label as an input and generates an adaptive acoustic model;
A post-adaptive acoustic model parameter memory for recording the adaptive acoustic model;
A second speech recognition processing unit that receives the speech digital signal and performs speech recognition processing with reference to the adaptive acoustic model recorded in the post-adaptation acoustic model parameter memory;
Comprising
The frame sound quality estimation unit calculates a GMM likelihood corresponding to the voice feature amount of the frame with reference to the GMM, and outputs the GMM likelihood as the frame sound quality.
The voice recognition processing control unit outputs a control signal including a non-recognition instruction signal indicating that voice recognition processing is not performed when the sound quality level is lower than a predetermined threshold,
The voice recognition processing unit does not perform voice recognition processing of the voice file when the control signal includes a non-recognition instruction signal.
A speech recognition apparatus characterized by that. The speech recognition apparatus according to any one of claims 1 to 4 ,
When the sound quality level is better than the predetermined threshold, the voice recognition processing control unit recognizes the voice with the beam search width being larger as the sound quality level is worse and the beam search width being smaller as the sound quality level is better. A speech recognition apparatus that outputs a control signal for setting a beam search width of processing. The speech recognition apparatus according to claim 3 ,
The audio file processing unit
If the sound quality level is lower than the minimum sound quality level for the acoustic model learning data , discard the audio file , or
A speech recognition apparatus, wherein a speech file is discarded when the value of the sound quality level is smaller than a threshold μ-2σ calculated from an average μ and a standard deviation σ of a sound quality level distribution of acoustic model learning data . A feature quantity analysis process in which a feature quantity analysis unit analyzes a voice feature quantity of a voice digital signal included in an input voice file in units of frames;
A frame sound quality estimation unit that evaluates the sound quality of the audio digital signal for each frame and outputs a frame sound quality that is the sound quality of the frame; and
An average sound quality estimation unit that calculates a sound quality level that is the sound quality of the sound file from the frame sound quality of all frames of the sound file;
A voice recognition process control unit for outputting a control signal for controlling an operation during the voice recognition process based on the sound quality level;
A speech recognition processing section in which the speech recognition processing unit performs speech recognition processing based on the control signal;
With
The frame sound quality estimation process calculates a GMM likelihood corresponding to the voice feature of the frame with reference to the GMM, and outputs the GMM likelihood as the frame sound quality.
The voice recognition processing control process outputs a control signal including a non-recognition instruction signal indicating that voice recognition processing is not performed when the sound quality level is lower than a predetermined threshold,
The voice recognition process does not perform voice recognition processing of the voice file when the control signal includes a non-recognition instruction signal.
A speech recognition method characterized by the above. The speech recognition method according to claim 7 ,
And an audio file control process in which an audio file control unit determines the audio file processing order in descending order of the audio quality level by inputting the audio file information of the audio digital signal, the audio quality level, and the control signal;
A voice file processing step in which the voice file processing unit records the voice digital signal in the voice file memory in units of the voice file and outputs the recorded voice digital signal to the voice recognition processing unit in the processing order;
A speech recognition method comprising: A feature quantity analysis process in which a feature quantity analysis unit analyzes a voice feature quantity of a voice digital signal included in an input voice file in units of frames;
A frame sound quality estimation unit that evaluates the sound quality of the audio digital signal for each frame and outputs a frame sound quality that is the sound quality of the frame; and
An average sound quality estimation unit that calculates a sound quality level that is the sound quality of the sound file from the frame sound quality of all frames of the sound file;
A voice recognition process control unit for outputting a control signal for controlling an operation during the voice recognition process based on the sound quality level;
A speech recognition processing step in which a speech recognition processing unit outputs the result of speech recognition processing based on the speech feature amount as an adaptation label by receiving the speech feature amount and the control signal;
An unsupervised adaptation unit learns an acoustic model using the adaptation label as an input and generates an adaptive acoustic model;
A second speech recognition processing section in which a second speech recognition processing unit receives the speech digital signal and performs speech recognition processing with reference to the adaptive acoustic model recorded in the post-adaptation acoustic model parameter memory;
With
The frame sound quality estimation process calculates a GMM likelihood corresponding to the voice feature of the frame with reference to the GMM, and outputs the GMM likelihood as the frame sound quality.
The voice recognition processing control process outputs a control signal including a non-recognition instruction signal indicating that voice recognition processing is not performed when the sound quality level is lower than a predetermined threshold,
The voice recognition process does not perform voice recognition processing of the voice file when the control signal includes a non-recognition instruction signal.
A speech recognition method characterized by the above. Device program for causing a computer to function as a speech recognition apparatus according to any one of claims 1 to 6. A computer-readable recording medium on which any of the apparatus programs according to claim 10 is recorded.

Images