JP3605308B2 - Voice recognition device and recording medium - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition device and a recording medium, and more particularly, to a speech recognition device and a recording device capable of reducing a load without lowering recognition accuracy in calculating a likelihood between an input acoustic feature and an acoustic feature dictionary. Regarding the medium.
[0002]
[Prior art]
FIG. 7 shows the configuration of a conventional general speech recognition apparatus. The speech input terminal 101, the acoustic feature calculator 102, the likelihood calculator 103, the acoustic feature dictionary 104, the recognition dictionary searcher 105, It comprises a recognition vocabulary dictionary 106 and a recognition result output terminal 107.
[0003]
The audio signal is input from the input terminal 101, is divided into frames of a predetermined time length (for example, 10 ms), and the feature quantity of the input speech is calculated by the acoustic feature quantity calculator 102 for each frame. As the feature amount used in the speech recognition, for example, a band filter bank output expressing a power spectrum shape, a cepstrum parameter, and the like are used.
[0004]
Likelihood calculator 103 calculates likelihood for each state of acoustic feature dictionary 104 using acoustic feature dictionary 104 created in advance for acoustic features of input speech calculated for each frame. .
Here, the likelihood is an index value indicating how similar the acoustic feature of the input voice is to the acoustic feature of each state of the acoustic feature dictionary.If the cepstrum is used as the acoustic feature, the input The cepstrum distance between the cepstrum of the speech and the cepstrum held for each state of the acoustic feature dictionary is used.
[0005]
The acoustic feature dictionary 104 divides a sound into predetermined states (units) and holds acoustic features for each state. Examples of states that are units of voice include phonemes (eg, 'a', 't') or triads of phonemes (eg, 'a;k;a') that take into account temporally preceding and succeeding phoneme environments. , 'T;a;k'), and a method using a unit obtained by temporally dividing one phoneme. The number of states varies depending on how the unit is selected. If a phoneme is used as a unit, the number of states becomes several hundreds if several tens or three groups of phonemes are used.
For example, the acoustic feature dictionary is composed of 256 elements. If the utterance is 1 second, the analysis is performed at an analysis period of 10 ms, and the acoustic feature of 100 frames is calculated. As a result, a 256 × 100 likelihood table is obtained. Is calculated.
[0006]
Next, when the utterance of the input speech is completed, the utterance content is searched by the recognition dictionary searcher 105 using the recognition vocabulary dictionary 106. The recognition vocabulary dictionary describes words to be recognized in units of recognition described above. For example, when "red" is expressed in units of the triplet phoneme exemplified in the above description, it is expressed as "-;a; k ',"a;k;a',"k;a;-". Here, "-" represents a silent state.
Hidden Markov models and Viterbi search techniques are used to search for the most likely words in the vocabulary to be recognized based on the time series of the likelihood calculated for each state. -Digital signal processing of sound information ", Shokodo, pp. 42-79).
[0007]
As described above, in the conventional speech recognition apparatus, the likelihood calculation is performed for all frames of the input signal, and therefore, there is a problem that the calculation amount is large. As a method for coping with this problem, Japanese Patent Laid-Open No. 2-239291 discloses a method of temporally examining the amount of change in the acoustic feature of a voice, and examining the frame at the time when the dynamic feature is large or at the maximum. There is disclosed a technique in which likelihood calculation is performed only in the case. This is intended to narrow down the number of candidates for the phoneme boundary position and to reduce the likelihood calculation amount.
[0008]
[Problems to be solved by the invention]
In order to obtain a high speech recognition rate, it is necessary to accurately analyze information such as plosives whose acoustic features change in a short period of time, and to divide the acoustic feature dictionary into a large number of states in consideration of the phoneme environment. It is desirable. If the analysis interval is shortened and the state of the acoustic feature dictionary is large, the likelihood calculation amount becomes extremely large, and the problem that the realization cost of the recognition device increases as described above occurs.
[0009]
The technique disclosed in Japanese Patent Application Laid-Open No. Hei 2-239291 proposed to solve this problem aims at detecting a punctuation between phonemes, and the likelihood of a frame having a large instantaneous change or a frame at a local maximum point. Therefore, there is a problem that the likelihood in a stable feature amount section cannot be calculated. Furthermore, depending on the voice, the amount of change is small or the maximum point does not appear, the change in the acoustic feature cannot be detected, and the likelihood cannot be obtained. As a result, in the above-described related art, there is a problem in that omitting the likelihood calculation in a section where the amount of change is small greatly deviates from the actual likelihood, leading to a reduction in speech recognition accuracy.
The present invention has been made in view of such a problem, and an object of the present invention is to reduce the load of likelihood calculation and increase the accuracy of speech recognition. It is to provide a recording medium.
[0012]
[Means for Solving the Problems]
The voice recognition device of the present invention divides a voice signal into frames of a predetermined time length, calculates a voice feature value, and converts the voice into a plurality of states based on a predetermined criterion. And an acoustic feature dictionary holding acoustic features for each of the classified states, and a likelihood calculator for calculating the acoustic features of the input speech and the likelihood for each state of the acoustic features dictionary. And a recognition vocabulary dictionary that describes speech recognition target words using the above-described state, and a recognition dictionary searcher that inputs a previously calculated likelihood calculation result and calculates a speech recognition result from the recognition vocabulary dictionary A change amount calculator for calculating a temporal change amount of an acoustic feature amount; a change amount memory for storing the change amount calculated by the change amount calculator over a plurality of frames; predetermined number of off from those variation greater from Likelihood calculation frame selector that selects the number of frames and performs likelihood calculation only on the selected frames, and controls the unselected frames to use the already calculated likelihood values. It is provided with.
[0013]
As a result, the maximum value of the calculation amount in the frame can be suppressed to a certain fixed amount, and a configuration suitable for a real-time processing device can be obtained.
Further, the likelihood calculator calculates and outputs the average value of the likelihood calculated immediately before and the likelihood calculated next for the section of the frame in which the likelihood calculation is omitted, so that the change amount is small. Thus, the likelihood can be obtained as an approximate solution while omitting the calculation of the likelihood.
Further, the likelihood calculator calculates and outputs the likelihood calculated immediately before and the slope value of the likelihood to be calculated next with respect to the section of the frame where the likelihood calculation is omitted, so that the change amount is small in the section. The likelihood can be obtained as an approximate solution while omitting the calculation of the likelihood.
[0014]
Further, the likelihood calculator outputs the result obtained by calculating the likelihood of the frame next to the frame of which the last likelihood calculation was performed as the likelihood of a subsequent frame period in which the likelihood calculation is not to be performed. It is possible to improve the accuracy of speech recognition using the likelihood that the amount has decreased.
Further, the likelihood calculator calculates the likelihood of a frame next to the last frame for which the likelihood calculation is performed last and a frame immediately before the frame for which the likelihood calculation is next performed. Alternatively, by calculating and outputting the inclination value, it becomes possible to increase the accuracy of speech recognition using the likelihood that the amount of change is reduced.
Further, the present invention is a computer-readable recording medium in which a program for causing a computer to function as the voice recognition device is recorded.
[0015]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. The components having the same functions as those in FIG. 7 are denoted by the same reference numerals.
FIG. 1 shows a first embodiment of the speech recognition apparatus of the present invention. In the circuit configuration of FIG. 7, newly added components are a change amount calculator 108 for calculating a change in the feature amount of the input signal, and a likelihood calculator based on the change amount of the feature amount of the input signal. Is a likelihood calculation controller 109 for controlling the operation of the above.
[0016]
The audio signal is input from the input terminal 101 and is divided into frames having a predetermined time length (for example, 10 ms). The feature quantity of the input speech is calculated by the acoustic feature quantity calculator 102 for each frame. For example, cepstrum is used as the feature amount. The acoustic feature amount calculated for each frame is input to the change amount calculator 108 and the likelihood calculator 103. The change amount calculator 108 aims to detect a change in the feature amount of the input audio signal, and calculates a change amount from, for example, a difference between the sound feature amount calculated in the immediately preceding frame and the sound feature amount in the current frame. When the state (for example, phoneme) of the input speech signal changes, the acoustic feature such as the spectrum shape changes, and the change in the feature is large. However, in a steady vowel section or the like, a relatively stable acoustic feature value continues, and the change amount of the acoustic feature value decreases.
[0017]
When the cepstrum is used as the acoustic feature amount, for example, the cepstrum distance between the cepstrum of the immediately preceding frame and the cepstrum of the current frame is used as the change amount. In addition, there is a method of using information of a temporal change of a feature amount generally referred to as a dynamic feature amount as an acoustic feature amount used in speech recognition (when a cepstrum is used as an acoustic feature, a feature called a delta cepstrum is used. There). In this case, the magnitude of the dynamic feature may be calculated as the change. The change amount of the feature amount calculated by the change amount calculator 108 is input to the likelihood calculation controller 109, and is compared with a preset determination threshold.
[0018]
As a result of comparing the change in the feature amount calculated by the change amount calculator 108 with the judgment threshold value of the likelihood calculation controller 109, the process branches when the change amount is larger than the judgment threshold value or smaller.
If the change is larger than the determination threshold, the state change of the input voice signal is large, and the likelihood calculator 103 is caused to execute the likelihood calculation for the feature of the input signal calculated by the acoustic feature calculator 102. The processing in this case is the same as the conventional one.
[0019]
Next, a part of the processing which is a feature of the present invention will be described. If the amount of change is smaller than the determination threshold, the state of the input speech signal has not changed much, and it is expected that the likelihood calculation result will not change significantly. In this case, the likelihood calculation controller 109 obtains the likelihood value based on a criterion different from the case where the amount of change is larger than the determination threshold. There may be one or more criteria.
[0020]
For example, a criterion that the likelihood calculator 103 does not perform the likelihood calculation of the current frame but uses the likelihood calculation result calculated immediately before is used. The voice signal includes a section in which the change of the acoustic feature amount is fast, such as a plosive consonant, and a section, such as a vowel stationary part, which does not change much. As described above, in a section where the change in the acoustic feature amount is small, the calculation amount can be reduced by controlling so that the likelihood calculation is not performed.
As described above, the likelihood calculation control in which the likelihood calculation is performed in the section where the change in the acoustic feature amount is large and the likelihood calculation is performed based on the criterion that the likelihood calculation is not performed in the section where the acoustic feature amount is small. Is shown in FIG.
[0021]
FIG. 6A shows the energy of the input audio signal, and FIG. 6B shows the magnitude of the change in the feature amount. The threshold indicated by the dotted line in FIG. 6B is a threshold for determining the amount of change. Likelihood calculation is performed for a frame whose amount of change is larger than this threshold, and the value of the likelihood calculated recently is used for a frame smaller than the threshold. . FIG. 6C shows the execution state of the likelihood calculation. The interval between the graduations represents the frame length, and the frame marked with c is a frame in which the amount of change in the feature amount is larger than the determination threshold and the likelihood calculation has been performed, and the frame not marked with c is , The likelihood value is a frame using the likelihood calculated immediately before.
[0022]
Subsequent processing is the same as that described in the prior art of FIG. 7, and when the utterance of the input speech is completed, the likelihood table obtained by omitting the likelihood calculation and the recognition vocabulary dictionary 106 are used. The utterance content is searched by the search device 105.
FIG. 2 shows a second embodiment of the speech recognition apparatus of the present invention. What is newly added as a component to FIG. 1 of the first embodiment is that the output of the likelihood calculation controller 109 is input, and the likelihood calculation is performed at a reduced frame interval in a section where the variation is small. Only the thinning-out calculation controller 201 that controls The parts different from the first embodiment will be described below.
[0023]
When a steady acoustic feature such as a vowel continues over a plurality of frames, the change in the acoustic feature of an adjacent frame is small, but in a plurality of frames, the amount of change accumulates and the amount of change increases. is there. In this case, if the likelihood calculation is not performed, the recognition rate decreases. Therefore, in the second embodiment, the thinning calculation controller 201 receives the output of the likelihood calculation controller 109 and, in a section where the amount of change is large, sends the output to the likelihood calculator 103 in the same manner as in the first embodiment. Execute likelihood calculation.
[0024]
On the other hand, in a section where the amount of change in the output of the likelihood calculation controller 109 is small, the number of processing frames is counted, and likelihood calculation is performed by the likelihood calculator 103 for a frame thinned out at a predetermined thinning interval. The control is performed so that the likelihood calculation is not performed in the other frames and the result of the likelihood calculation calculated immediately before is used.
The state of likelihood calculation control is shown in FIG. In this example, the control is performed with the thinning rate of 2, so that the likelihood calculation is performed once every two frames even in a section where the amount of change in the feature amount is smaller than the determination threshold as compared with FIG. Has become. Note that the thinning rate does not need to be limited to 2 and is optional.
[0025]
FIG. 3 shows a third embodiment of the speech recognition apparatus of the present invention. A new component added to FIG. 1 is a change amount memory 301 for storing the change amounts of the feature amounts of a plurality of frames, and a frame for performing likelihood calculation based on a certain selection criterion is selected from the change amount memory 301. Is the likelihood calculation frame selector 302. Hereinafter, portions different from the first and second embodiments will be described.
In the first and second embodiments, the likelihood calculation is performed only on a frame that satisfies a certain criterion, instead of all input frames, and the amount of calculation as a whole is reduced. However, in an apparatus that performs real-time processing, it is necessary to keep the amount of calculation constant in predetermined processing units.
[0026]
In the third embodiment, in order to address this problem, processing is performed for each of a predetermined number (M) of frames, and a frame (N, where N, <M) and control to use the recently calculated likelihood values for the remaining frames.
As an example of the criterion for this selection, the change amount calculated by the change amount calculator 108 is stored in a change amount memory 301 over M frames, and the likelihood calculation frame selector 302 selects the M change amounts. To select N frames from the largest one.
By performing such control, the maximum value of the amount of calculation in the M frames can be suppressed to a certain fixed amount, and a configuration suitable for a real-time processing device can be obtained.
[0027]
FIG. 4 shows a fourth embodiment of the speech recognition apparatus of the present invention. The average likelihood calculator 601 in this figure substitutes for the likelihood calculator 103 in FIGS. For a frame in which the amount of change in the feature of the input speech exceeds the threshold, the likelihood calculation is performed as before, and the calculation result is stored in a buffer memory in the average likelihood calculator 601. Assuming that the value of the immediately preceding likelihood obtained in this manner is L, and then the likelihood N calculated with the amount of change exceeding the threshold, the frame between the frame where L is calculated and the frame where N is calculated is defined as On the other hand, the likelihood of (L + N) / 2 is applied. Other processes are in accordance with the first to third embodiments.
FIG. 6E shows the result of the likelihood calculation control. This result is applied to FIG. 6C, and the calculation of the likelihood is the same as that of FIG. 6C. The frame for which the likelihood calculation is omitted has the average value of the likelihoods calculated at both ends thereof (only the first likelihood non-calculation section is shown in the figure to avoid complication).
[0028]
FIG. 5 shows a speech recognition apparatus according to a fifth embodiment of the present invention. The inclination likelihood calculator 701 in this figure substitutes for the likelihood calculator 103 in FIGS. For a frame in which the amount of change in the feature of the input speech exceeds the threshold, the likelihood calculation is performed in the conventional manner, and the result of the calculation is stored in the buffer memory in the gradient likelihood calculator 701. Assuming that the immediately preceding likelihood value obtained in this way is L, and then the likelihood N calculated with the change amount exceeding the threshold, the frame between the frame where L was calculated and the frame where N was calculated was calculated. The number P is counted, and a slope likelihood of L + (NL) × m / (P + 1) (m is an integer from 1 to P) is applied to these P frames. Other processes are in accordance with the first to third embodiments.
[0029]
FIG. 6F shows the result of the likelihood calculation control. This result is applied to FIG. 6C, and the calculation of the likelihood is the same as that of FIG. 6C. Frames for which the likelihood calculation has been omitted have the likelihood gradient distribution values obtained at both ends thereof (only the first likelihood non-calculation section is shown in the figure to avoid complication).
In addition to the above-described first to fifth embodiments, the following embodiments are also possible.
[0030]
In FIG. 2, in the case of frames in which the acoustic feature amount does not exceed the threshold, in the first embodiment, a value obtained by finally calculating the likelihood is used as the likelihood of these frames. Instead of the last calculated likelihood as the likelihood of a frame whose acoustic feature value does not exceed the threshold value, the likelihood of the next frame lastly calculated may be calculated and this value may be used. . As a result, the accuracy of speech recognition can be improved by using the likelihood of a small change in the amount of feature, instead of the likelihood of a large change in the amount of feature, as the likelihood of a frame whose acoustic feature does not exceed the threshold. It becomes possible. The result of the likelihood calculation control obtained in this way is shown in FIG.
[0031]
Furthermore, in FIG. 4 or FIG. 5, in the case of the frame in which the acoustic feature amount does not exceed the threshold value, in the embodiment of FIG. Although the likelihood is calculated also in the frame of the above, the likelihood is calculated also in the frame in which the change amount exceeds the threshold value and the change amount immediately before the likelihood calculation does not exceed the threshold value. The average value or the slope value of the likelihood calculated at both ends of the frame not exceeding the threshold value is applied to the frame during this period. Note that the average likelihood value or the slope likelihood value is based on the method described in the fourth or fifth embodiment. FIG. 6H shows the result of the likelihood calculation control thus obtained.
[0032]
Note that the present invention is not limited to the above embodiment.
In each of the above embodiments, a single likelihood calculation control criterion is used for a section in which a change in the feature amount of the input speech is small. However, the present invention can be implemented by combining a plurality of criteria, for example.
Further, the present invention may be a computer-readable recording medium on which a program for causing a computer to function as the voice recognition device is recorded. For example, any of a magnetic tape, a CD-ROM, an IC card, a RAM card, etc. It may be a type of recording medium.
[0033]
【The invention's effect】
As described above in detail, according to the present invention, the change amount calculator that calculates the temporal change amount of the acoustic feature amount, the calculated change amount is compared with a predetermined threshold value, and the threshold value is exceeded. The likelihood calculation controller that calculates and outputs the likelihood value using a plurality of criteria in the case and the case where the likelihood value does not exceed is provided. The likelihood calculation is executed on the other hand, and when the change is small, for example, control is performed so as to use the latest calculated likelihood value, or control such as thinning out the likelihood calculation becomes possible. Deterioration can be kept small, and the amount of calculation can be reduced.
[0034]
In addition, not only is the likelihood calculated for a section where the change in the feature amount of the input speech is large, but also an approximate likelihood can be assigned to a section where the change in the feature amount of the input voice is small or a stable section. Consistent input speech recognition is possible. In addition, since the likelihood can be obtained regardless of the magnitude of the change in the feature amount of the input speech, the likelihood can be obtained even when the variation is small or the maximum point does not appear depending on the speech. Therefore, according to the present invention, the likelihood can be obtained as an approximate solution while omitting the likelihood calculation even in a section where the amount of change is small, so that the calculation amount can be reduced without lowering the speech recognition accuracy.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a first embodiment of a speech recognition device of the present invention.
FIG. 2 is a block diagram showing a second embodiment of the speech recognition apparatus of the present invention.
FIG. 3 is a block diagram showing a third embodiment of the speech recognition device of the present invention.
FIG. 4 is a block diagram showing a fourth embodiment of the speech recognition device of the present invention.
FIG. 5 is a block diagram showing a fifth embodiment of the speech recognition apparatus of the present invention.
FIGS. 6A to 6H are explanatory diagrams of likelihood calculation control.
FIG. 7 is a view for explaining an embodiment of a conventional speech recognition apparatus.
[Explanation of symbols]
101 input terminal 102 acoustic feature calculator 103 likelihood calculator 104 acoustic feature dictionary 105 recognition dictionary searcher 106 recognition vocabulary dictionary 107 recognition result output terminal 108 change calculator 109 likelihood calculation controller 201 thinning calculation controller 301 Change amount memory 302 Likelihood calculation frame selector 601 Average likelihood calculator 701 Slope likelihood calculator

Claims (6)

The audio signal is divided into frames of a predetermined time length, and an acoustic feature amount calculator that calculates a feature amount of the audio, and the audio is classified into a plurality of states according to a predetermined criterion, and the classified state is An acoustic feature dictionary holding acoustic features for each speech, an acoustic feature of an input speech and a likelihood calculator for calculating likelihood for each state of the acoustic feature dictionary, and a speech recognition target word as described above. A speech recognition device comprising a recognition vocabulary dictionary described using the state of a change amount calculator for calculating a temporal change amount of the characteristic amount, a change amount memory for storing the amount of change calculated by the change amount calculator over a plurality of frames, the change from the plurality of frames predetermined number of off from those amounts large Likelihood calculation frame selector that selects the number of frames, performs likelihood calculation only on the selected frame, and controls the non-selected frame to use the already calculated likelihood value. A speech recognition device comprising: The said likelihood calculator calculates and outputs the average value of the likelihood calculated immediately before and the likelihood calculated next with respect to the area of the frame from which the likelihood calculation was omitted. Voice recognition device. 2. The likelihood calculator according to claim 1, wherein the likelihood calculator calculates and outputs, for a section of the frame in which the likelihood calculation is omitted, the likelihood calculated immediately before and the slope value of the likelihood calculated next. 3. Voice recognition device. The said likelihood calculator outputs the result obtained by calculating the likelihood of the frame next to the frame of which the last likelihood calculation was performed as the likelihood of the subsequent frame period in which the likelihood calculation is not to be performed. Item 2. The speech recognition device according to item 1. The likelihood calculator calculates the likelihood of the frame immediately following the frame next to the last frame for which the likelihood calculation is performed and the frame immediately before the frame for which the likelihood calculation is next performed. The speech recognition device according to claim 1, wherein a value is calculated and output. A computer-readable recording medium on which a program for causing a computer to function as the speech recognition device according to claim 1 is recorded.

Images