JP3584002B2 - Voice recognition device and voice recognition method - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition device and a speech recognition method, and more particularly to a method for accurately determining an appropriate word from speech including unnecessary words.
[0002]
[Prior art]
As a method of recognizing a word in a word dictionary from a voice including an unnecessary word (for example, a meaningless voice or a particle), for example, a method described in Japanese Patent Application Laid-Open No. 7-77998 is known.
[0003]
In such a conventional speech recognition method, first, a feature amount of an unnecessary word is generated from speech feature amounts of all recognition candidates (necessary words) stored in a word dictionary, and the feature amount of the unnecessary word is stored in a recognition dictionary in advance. Register. Then, of the input speech signals, those that match the feature amount of the unnecessary word are recognized as unnecessary words, and the words recognized as unnecessary words are removed from the recognition result to obtain the voice recognition result of the necessary word. It is.
[0004]
Generation of the feature amount of the unnecessary word in such a conventional method is performed as follows. First, prior to generation of the feature amount of the unnecessary word, the feature amount of the necessary word is generated. The feature amount of the necessary word is generated by inputting several types of voice feature amounts as samples for one necessary word, and acoustically analyzing and learning each of the samples to generate the feature amount of the necessary word. Next, the feature amounts of all the necessary words thus generated are averaged, and the average value is set as the feature amount of the unnecessary word.
[0005]
[Problems to be solved by the invention]
As described above, in the above-described conventional recognition method, the speech feature amount of the unnecessary word is generated from the speech feature amounts of all the recognition candidates (necessary words). Therefore, the recognition candidates in the recognition dictionary are added and corrected. Each time, it is necessary to regenerate the speech feature amount of the unnecessary word one by one, so that a troublesome work is required in adding or deleting a recognition candidate.
[0006]
Also, at the time of voice recognition processing, likelihood calculation and approximate distance calculation with the input voice signal must be performed not only for the necessary words but also for the voice feature amounts of the unnecessary words, and a voice recognition processing step is added accordingly. Therefore, there is a problem that the time required for deriving the recognition result becomes long.
[0007]
Therefore, an object of the present invention is to provide a speech recognition method that can quickly perform speech recognition even when a recognition candidate is added or deleted.
[0008]
[Means for Solving the Problems]
In view of the above problems, the present invention has the following features.
[0009]
The invention according to claim 1 relates to a speech recognition apparatus, comprising: an acoustic analysis unit for acoustically analyzing an input speech signal to extract a frame feature; and a word model by connecting a silence model to both ends of a word reference model. Word model generating means for generating the likelihood calculation means; comparing the frame feature quantity and the feature quantity of the reference model to calculate the likelihood of the frame feature quantity with respect to the word model; and the calculated likelihood. Matching operation means for calculating the degree of matching of the word model with respect to the input speech signal based on the above, recognition candidate setting means for setting recognition candidates in accordance with the degree of matching, and garbage setting the garbage likelihood for the silence model Having likelihood setting means,The garbage likelihood setting means calculates the garbage likelihood by averaging the likelihoods of the frame features from the top model to the Nth among the likelihoods with respect to the reference model feature,The matching operationmeansIs characterized in that a matching operation is performed using the garbage likelihood set by the garbage likelihood setting means as the likelihood of the silent model.
The invention according to claim 2 relates to a speech recognition apparatus, comprising: an acoustic analysis means for acoustically analyzing an input speech signal to extract a frame feature; and a word model by connecting a silence model to both ends of a word reference model. Word model generating means for generating the likelihood calculation means; comparing the frame feature quantity and the feature quantity of the reference model to calculate the likelihood of the frame feature quantity with respect to the word model; and the calculated likelihood. Matching operation means for calculating the degree of matching of the word model with respect to the input speech signal based on the above, recognition candidate setting means for setting recognition candidates in accordance with the degree of matching, and garbage setting the garbage likelihood for the silence model Having likelihood setting means,The garbage likelihood setting means calculates the garbage likelihood by averaging the likelihoods of the frame features from the top model to the Nth among the likelihoods with respect to the reference model feature,The matching operationmeansIs the likelihood calculation as the likelihood of the silence model.meansAnd a garbage likelihood set by the garbage likelihood setting means is selected to perform a matching calculation.
[0010]
The invention of claim 3 isThe present invention relates to a speech recognition apparatus, and relates to acoustic analysis means for acoustically analyzing an input speech signal to extract a frame feature, and word model creation means for creating a word model by connecting a silence model to both ends of a word reference model. And a likelihood calculating means for comparing the frame feature quantity with the feature quantity of the reference model to calculate the likelihood of the frame feature quantity with respect to the word model, and a likelihood calculating means for calculating the likelihood of the word model based on the calculated likelihood. Matching calculation means for calculating the degree of matching with respect to the input voice signal, recognition candidate setting means for setting recognition candidates according to the degree of matching, and garbage likelihood setting means for setting garbage likelihood for the silent model. The garbage likelihood setting means determines whether the magnitude of the likelihood is higher in the likelihood of the frame feature with respect to the reference model feature. The K-th likelihood and garbage likelihood, the matching calculation means, as the likelihood of silence model, and performing a matching operation using the garbage likelihood set by the garbage likelihood setting means.
[0011]
The invention of claim 4 isThe present invention relates to a speech recognition apparatus, and relates to acoustic analysis means for acoustically analyzing an input speech signal to extract a frame feature, and word model creation means for creating a word model by connecting a silence model to both ends of a word reference model. And a likelihood calculating means for comparing the frame feature quantity with the feature quantity of the reference model to calculate the likelihood of the frame feature quantity with respect to the word model, and a likelihood calculating means for calculating the likelihood of the word model based on the calculated likelihood. Matching calculation means for calculating the degree of matching with respect to the input voice signal, recognition candidate setting means for setting recognition candidates according to the degree of matching, and garbage likelihood setting means for setting garbage likelihood for the silent model. The garbage likelihood setting means determines whether the magnitude of the likelihood is higher in the likelihood of the frame feature with respect to the reference model feature. The K-th likelihood is set as the garbage likelihood, and the matching calculation means is set as the likelihood of the silence model, the likelihood of the silence model calculated by the likelihood calculation means, and the garbage likelihood setting means. It is characterized in that one of the garbage likelihoods is selected and a matching operation is performed..
[0012]
The invention of claim 5 isIn addition to the features of claim 2 or 4, furthermore, the matching calculation means compares the likelihood of the silence model calculated by the likelihood calculation means with the garbage likelihood set by the garbage likelihood setting means, The feature is that any one of the larger likelihoods is selected.
[0013]
The invention of claim 6 isThe present invention relates to a speech recognition method, comprising the steps of: acoustically analyzing an input speech signal to extract a frame feature amount; connecting a silence model to both ends of a reference model of a word to create a word model; Calculating the likelihood of the frame feature for the word model by comparing the amount with the feature of the reference model; and determining the degree of matching of the word model to the input speech signal based on the calculated likelihood. Computing, setting a recognition candidate according to the degree of matching, and setting a garbage likelihood for the silence model, wherein the garbage likelihood setting step comprises: The garbage likelihood is calculated by averaging the likelihoods from the top to the Nth likelihood of the likelihood for the feature amount. Calculating a step of the matching operation, as the likelihood of silence model, and performing a matching operation using the garbage likelihood set by the garbage likelihood setting step.
[0014]
The invention of claim 7 isThe present invention relates to a speech recognition method, comprising the steps of: acoustically analyzing an input speech signal to extract a frame feature amount; connecting a silence model to both ends of a reference model of a word to create a word model; Calculating the likelihood of the frame feature for the word model by comparing the amount with the feature of the reference model; and determining the degree of matching of the word model to the input speech signal based on the calculated likelihood. Computing, setting a recognition candidate according to the degree of matching, and setting a garbage likelihood for the silence model, wherein the garbage likelihood setting step comprises: The garbage likelihood is calculated by averaging the likelihoods from the top to the Nth likelihood of the likelihood for the feature amount. And the matching calculation step is any of the silence model likelihood calculated in the likelihood calculation step and the garbage likelihood set in the garbage likelihood setting step, as the likelihood of the silence model. A matching operation is performed by selecting one of them.
[0015]
The invention of claim 8 relates to a speech recognition method, wherein a step of acoustically analyzing an input speech signal to extract a frame feature amount, and a step of connecting a silence model to both ends of a reference model of a word to create a word model Calculating the likelihood of the frame feature with respect to the word model by comparing the frame feature with the feature of the reference model, and calculating the likelihood of the frame model with respect to the word model based on the calculated likelihood. Calculating a matching degree for the input voice signal, setting a recognition candidate according to the matching degree, and setting a garbage likelihood for the silence model,In the garbage likelihood setting step, among the likelihoods of the frame feature amount with respect to the reference model feature amount, the K-th likelihood having the highest likelihood as the garbage likelihood is defined as garbage likelihood,In the matching calculation step, the matching calculation is performed using the garbage likelihood set in the garbage likelihood setting step as the likelihood of the silence model.
A ninth aspect of the present invention relates to a speech recognition method, wherein a step of acoustically analyzing an input speech signal to extract a frame feature amount, and a step of connecting a silence model to both ends of a word reference model to create a word model Calculating the likelihood of the frame feature with respect to the word model by comparing the frame feature with the feature of the reference model, and calculating the likelihood of the frame model with respect to the word model based on the calculated likelihood. Calculating a matching degree for the input voice signal, setting a recognition candidate according to the matching degree, and setting a garbage likelihood for the silence model,In the garbage likelihood setting step, among the likelihoods of the frame feature amount with respect to the reference model feature amount, the K-th likelihood having the highest likelihood as the garbage likelihood is defined as garbage likelihood,The matching calculation step includes, as the likelihood of the silence model, the likelihood of the silence model, the likelihood of the silence model calculated by the likelihood calculation step, and the garbage likelihood set by the garbage likelihood setting step. The matching operation is performed by selecting any one of the above.
[0016]
The invention of claim 10 isClaim 7 orIn addition to the above-mentioned features, the matching calculation step further includes the likelihood of the silence model calculated in the likelihood calculation step and the garbage likelihood setting.StepsThe garbage likelihood set by the above is compared, and the larger likelihood is selected.
[0021]
The features of the present invention and the effects thereof will be apparent by referring to the following embodiments.
[0022]
BEST MODE FOR CARRYING OUT THE INVENTION
First, an outline of a character recognition device and a character recognition method according to the present embodiment will be described with reference to FIGS.
[0023]
When a voice is input to the voice recognition device, a voice signal to be recognized is cut out from the voice input signal. This extraction is performed, for example, by monitoring the power of the audio input signal.
[0024]
That is, when a voice is input to the microphone, the power of the input voice signal from the microphone rises from the silence level. The section from this rising until the next level of the audio signal reaches the silence level is the section of the audio signal that should be originally recognized. If such an audio signal section is cut out as it is, the beginning or end of the audio signal to be recognized may be cut off. Therefore, in order to ensure that the voice signal section is to be recognized, the voice signal is usually cut out so as to include only a certain section of the silent signal before and after the voice signal section.
[0025]
For example, in FIG. 1 to FIG. 3, the cut-out audio signal is “XX”. Here, ○ is a silent signal. The meaning of the language of the audio signal actually input to the microphone is “d, the image is large”. Among them, “e-” is a meaningless unnecessary word that is often uttered by the input person when starting the voice input. “Image” is a recognition candidate (word) registered in the recognition dictionary, “GA” is a particle (unnecessary word) not in the recognition dictionary, and “large” is a recognition candidate (word) registered in the recognition dictionary.
[0026]
The audio signal input and cut out as described above is subjected to acoustic analysis at regular intervals (frame periods), and acoustic feature values (hereinafter, referred to as “frame feature values”) are extracted. Each extracted frame feature value is compared with the feature value of the reference syllable, and the degree of approximation with each reference syllable, that is, the likelihood is calculated.
[0027]
Here, the memory of the voice recognition device stores the feature values of the reference syllables, that is, 50 sounds (aiueo, etc.), voiced sounds (gakiguge, etc.), semi-voiced sounds (ぱ ぴ ぷ ぺ ぽ, etc.), muddy sounds (gyugyugyo, ) Are stored in advance together with the feature amounts of the respective syllables.
[0028]
The frame features extracted by acoustic analysis at the frame period are compared with the features of all the reference syllables stored in the memory, and the likelihood is calculated for each reference syllable. For example, in FIG. 1, the frame feature amounts are “O”, “O”, “E”, “E”, “GA”, “ZO”, “U”, “GA”, “O”, “O”. , "Ki", "i", "o", and "o", the feature amounts of these frame and the reference syllable are sequentially compared to calculate the likelihood.
[0029]
In the block shown in the lower part of FIG. 1, the leftmost column is the above-mentioned reference syllable, and the numbers in the rightmost columns following the leftmost column are the likelihood of each frame feature amount for each of these reference syllables. It is. For example, in FIG. 1, among audio signals to be recognized (audio signals cut out so as to include silence before and after), the first frame feature amount corresponds to a portion corresponding to a part of “○ (silence signal)”. Then, this feature amount is sequentially compared with all the feature amounts of each reference syllable, and the likelihood for each reference syllable is calculated. In FIG. 1, the likelihood for the reference syllable “a” is 0.1, the likelihood for the reference syllable “i” is 0.1,..., And the likelihood for the reference syllable “silence” is 0.9.
[0030]
When the calculation of the likelihood for the first frame feature is completed in this way, the likelihood for the reference syllable is calculated for the next frame feature “「 (silence signal) ”. Hereinafter, similarly, the reference syllables “A”, “A”, “E”, “E”, “GA”,... , "う",..., "ぽ", "ぽ", "○ (silence)" are calculated.
[0031]
The likelihood calculated in this manner is written on a memory (RAM: Random Access Memory) so that each likelihood is mapped on a matrix having a reference syllable and each frame as a correlation axis. . That is, the likelihood on the matrix shown in the lower part of FIG. 1 is directly mapped and stored in the memory.
[0032]
As described above, when the likelihood for each reference syllable is calculated and written into the memory, the degree of matching of the speech input signal to one recognition candidate in the word dictionary is calculated.
[0033]
The upper block in FIG. 1 shows the score (matching degree) of the likelihood for the recognition candidate “large”.
[0034]
As described above, since the speech signal is cut out so as to include a silent part, when calculating the matching of the speech signal with the recognition candidate, the recognition candidate syllable is obtained by adding a silent syllable before and after the recognition candidate. That is, as shown in FIG. 1, if the recognition candidate is “big”, the syllable “O” (silent) is added before and after each of the syllables “o”, “o”, “ki”, and “i” that constitute “big”. )], And a concatenation of these syllables is taken as a recognition candidate syllable.
[0035]
When the configuration of the recognition candidate syllable is made in this way, the likelihood of the frame feature of the speech signal is allocated to each syllable of the recognition candidate syllable.
[0036]
First, the likelihood between the frame feature first extracted from the audio signal (the feature of the “○ (silence signal)” portion) and each syllable of the candidate syllable is read from the RAM. That is, of the likelihood of the audio signal with respect to the reference syllable stored in the RAM (the likelihood allocated to the lower block in FIG. 1), each syllable “O” corresponding to the frame feature amount “O (silence signal)” is used. (Silence) ”,“ O ”,“ O ”,“ K ”,“ I ”,“ O (silence) ”are read out from the RAM, and these likelihoods are identified in the upper block in FIG. One frame feature amount and each syllable of these recognition candidates are allocated in a column where they intersect.
[0037]
Next, the likelihood between the frame feature quantity (the feature quantity of the “○ (silent signal)” portion) extracted second from the audio signal and each syllable is read from the RAM, and this is read in the same manner as above. Then, it is distributed in the column of the upper block in FIG.
[0038]
Hereinafter, similarly, the likelihood from the third extracted frame feature amount to the last extracted feature amount is assigned to the upper block in FIG.
[0039]
Actually, the likelihood assigned to the upper block in FIG. 1 is stored in a predetermined area in the RAM, using each frame and the recognition candidate syllable as a correlation axis.
[0040]
When the setting and distribution of the likelihood of the speech signal to the recognition candidate syllable are completed as described above, the degree of matching of the speech signal to the recognition syllable is calculated using the likelihood group distributed in this manner. Is done.
[0041]
In order to calculate the matching degree, first, a route that proceeds through each column from the lower left corner column to the upper right corner column in FIG. 1 is set, and the total value of likelihood in each column on the route is calculated. . Such a route setting is set, for example, such that the preceding column when viewed from one column is either left side or diagonally lower left. Alternatively, instead of this, the route may be set such that the preceding column when viewed from one column is any of the left side, the diagonally lower left, or the right below. There are various formulations for such DP matching, and for example, those described in Tokai University Press “Digital Audio Processing (1st printing)”, pages 167 to P167, can be used.
[0042]
In the present embodiment, the description is made using the likelihood. However, this may be log likelihood or a value based on the reciprocal of the distance (the difference between the frame feature and the feature of each syllable: absolute value). But it's fine. If a transition probability and the like are added to this, the well-known HMM (Hidden Markov Model) can be used.
[0043]
The total value of the likelihoods is calculated for all the routes that can be set in this way, and the largest total value among the total values calculated for each route is matched with the speech signal for the recognition candidate. Degree (score).
[0044]
In the example of FIG. 1, the recognition candidate is “big” and the voice signal includes a “big” portion. Therefore, in the upper block of FIG. And the recognition candidate syllable have a high likelihood in the column where they intersect. Therefore, the degree of matching (score) for the recognition candidate “large” becomes large due to the influence of the likelihood in the intersecting columns.
[0045]
On the other hand, when the matching degree (score) is calculated for a recognition candidate not included in the audio signal (for example, “small”), the likelihood of each column in the upper block in FIG. Therefore, the degree of matching (score) is also low.
[0046]
Therefore, the degree of matching (score) calculated for each recognition candidate is compared with each other, and the top several recognition candidates are selected in order from the one with the highest score and set as provisional recognition candidates. It is highly likely that proper recognition candidates are included.
[0047]
Then, for example, all of the provisional recognition candidates are displayed on the monitor of the voice recognition device, and the operator is allowed to select an appropriate one from among them, thereby to determine the recognition result.
[0048]
Alternatively, instead of such a method, recognition candidates may be classified in advance by content type (for example, classification of size such as “large” or “small”, classification of information such as “image” or “sound”, etc.). Various dictionaries are constructed, and the top several (e.g., five) recognition candidates having the matching degree (score) are extracted from each dictionary, and a combination of the recognition candidates from each extracted dictionary is created ( For example, if there are five recognition candidates in two dictionaries, 5 × 5 = 25 combinations), the combined recognition candidates are connected silently, and silence is added before and after to generate recognition candidate syllables. (For example, in the case of a combination of “gaze” and “big”, the recognition candidate syllable of “XX gaze XX big XX”), the matching degree between this recognition candidate syllable and the above speech signal (score The may be calculated again in the same manner as described above, is determined as a recognition result having the highest score among the combinations.
[0049]
As described above, a silence syllable is added before and after a recognition candidate to generate a recognition candidate syllable, and if the degree of matching between the recognition candidate syllable and the audio signal is determined, the silence signal portion is determined as described above. Even if the speech signal is cut out, the effect of the silent signal portion is absorbed by the recognition candidate syllable “○ (silence)” added before and after the recognition candidate, so that a relatively accurate recognition result can be obtained. Will be able to
[0050]
However, when an unnecessary word is added to the voice signal, the feature amount of the recognition candidate syllable “○ (silence)” and the feature amount of the unnecessary word are usually non-approximate, and thus the recognition candidate syllable “「 ( The effect of the unnecessary word cannot be absorbed by "silence"), and therefore, the accuracy of determining the recognition candidate is reduced.
[0051]
Therefore, in the present embodiment, the method of assigning the likelihood to the recognition candidate syllable “○ (silence)” is improved, thereby not only the effect of the silence part included in the audio signal but also the unnecessary words in the audio signal. The effect can be absorbed at the same time. More specifically, the likelihood of each frame feature extracted by acoustically analyzing the cut-out audio signal at a frame cycle with respect to the feature of the recognition candidate syllable “○ (silence)” (hereinafter referred to as “silence model likelihood”) Improve the settings of.
[0052]
More specifically, as shown in FIG. 2, the highest likelihood of the likelihood of each frame feature for each standard syllable is set as the silence model likelihood. That is, in the upper block of FIG. 2, the silence model likelihood of the frame feature amounts “e”, “e”, “ga”, “zo”, “u”, “ga”, “o”, “o”, “ki” and “i” are: 0.9, 0.9, 0.9, 0.9, 0.9, 0.8, 0.9, and 1.0, which are the maximum values of the likelihood of these frame feature amounts with respect to the reference syllable, are assigned, respectively. .
[0053]
When the likelihood is assigned in this manner, the route having the highest matching degree score passes through the silence model likelihood column in the upper block in FIG. The route proceeds diagonally to the upper right at the portions of "o", "ki" and "i", and then becomes the route again passing through the silence model likelihood column. That is, in the cut-out audio signal, the part of “○”, “○”, “ga”, “zo”, “u”, “ga” and the part of “○” and “○” following “big” Since the likelihood is set to the highest value, the route having the highest likelihood value usually passes through the column of the silence model likelihood.
[0054]
Therefore, even if the cut-out voice includes unnecessary words in addition to the silent signal portion, all the likelihood disturbances due to the silent signal portion and the unnecessary words are absorbed by the silent model likelihood. It will be.
[0055]
By the way, in the above embodiment, the maximum value of the likelihood of each frame feature amount with respect to the reference syllable is set as the silence model likelihood. There is a disadvantage that the likelihood of the portion to be emphasized, that is, the portion corresponding to the recognition candidate is not emphasized.
[0056]
For example, in the upper block of FIG. 2, the likelihood in the column on the route indicated by the arrow is a portion where the recognition candidate and the speech signal match, and thus the likelihood is sufficiently emphasized compared to the likelihood in the other columns. Must have been. However, in the column above such an arrow, as described above, the same likelihood as the silence model likelihood of the section is set. For this reason, the likelihood in the column on the arrow which should originally greatly affect the matching degree (score) is not emphasized so much, and as a result, the accuracy of the recognition result is likely to be affected by disturbance or the like. Disadvantages occur.
[0057]
Therefore, in order to improve such inconvenience, in the embodiment of FIG. 3, among the likelihoods of each frame feature quantity with respect to the reference syllable, the average value of the top N likelihoods (hereinafter, “garbage likelihood”) Is calculated, and when the garbage likelihood is larger than the silence model likelihood, the garbage likelihood is replaced with the silence model likelihood.
[0058]
When the garbage likelihood is replaced in this way, as shown in FIG. 3, the likelihood of the column on the arrow route becomes several steps larger than the silence model likelihood, and therefore, the arrow which should be emphasized originally The likelihood in the column on the route is effectively emphasized. In addition, the silence model likelihood of the silence portion of the speech signal is appropriately emphasized, and the silence model likelihood of the unnecessary word portion (including the “gaze” portion that is not the recognition target) is also properly emphasized. In addition, the effect of the silence part and the unnecessary word part can be effectively absorbed by the silence model likelihood of the period.
[0059]
The above is the outline of the present embodiment. Hereinafter, various examples illustrating the present embodiment in further detail will be described.
[0060]
FIG. 4 shows a block diagram of this embodiment. In the figure, reference numeral 1 denotes an audio input unit such as a microphone, and 2 denotes an audio signal cutout unit that cuts out an audio signal from an audio input signal from the audio input unit. As described above, the audio signal cutout unit 2 monitors the power of the audio input signal, and cuts out the audio signal so as to include a silent signal before and after.
[0061]
Reference numeral 3 denotes an acoustic analysis unit that performs acoustic analysis on the cut-out audio signal at predetermined frame periods, and extracts characteristic parameters (hereinafter, referred to as “frame characteristic parameters”). The frame feature parameters include, for example, a linear prediction coefficient, an LPC cepstrum, and energy for each frequency band. Since such acoustic analysis is already well known, a detailed description is omitted here. Note that such a frame feature parameter is synonymous with the frame feature amount in the above embodiment.
[0062]
Reference numeral 4 denotes a reference model parameter section that stores acoustic characteristic parameters for each reference model. The reference model parameter section 4 performs acoustic analysis of the reference model by the same method as the above-described acoustic analysis section 3 and stores the parameters as reference parameters of each model. Here, the reference model corresponds to, for example, a reference syllable in the above embodiment. Such a reference model includes a silent model as described in the above embodiment. Such a reference model may be a reference syllable as in the above embodiment, or may be a reference phoneme instead. In addition, the feature parameters of each whole word can be used as a reference model. Note that Hidden Markov Model or the like can be used as the reference model. Further, the reference model parameters can be represented by a discrete distribution, a continuous distribution, or the like.
[0063]
Reference numeral 5 denotes a likelihood calculation unit which compares the frame feature parameter for each predetermined frame period extracted by the acoustic analysis unit 3 with the feature parameter for each reference model of the reference parameter unit 4, and calculates the likelihood between the two. . As a method of calculating the likelihood, for example, a well-known method described in Chapter 3 of “Speech Recognition by Stochastic Model” issued by the Institute of Electronics, Information and Communication Engineers can be used.
[0064]
A RAM unit 6 stores the likelihood for each frame calculated by the likelihood calculating unit 5 in association with each reference model. For example, the likelihoods shown in the form of a matrix in the lower part of FIGS. 1 to 3 shown in the above embodiment are mapped on the RAM and stored.
[0065]
Reference numeral 7 denotes a recognition dictionary, which stores words as recognition candidates. As shown in FIG. 5, a plurality of recognition dictionaries are prepared in the recognition dictionary section, which are divided into categories such as "size" and "type of information".
[0066]
Reference numeral 8 denotes a word model creation unit that connects a reference model of a word to be recognized and evaluated and adds a silence model before and after the reference model to create a word model. For example, as shown in FIG. 6, if the word to be recognized is “big”, the created word model is “と big ○” (○ is a silent model).
[0067]
Reference numeral 9 denotes a matching calculation unit which refers to the likelihood for each reference model stored in the RAM unit 6 as described above and the garbage likelihood from a garbage likelihood calculation unit 10 described later, and The degree of matching (score) is calculated for the word model. The calculation of the matching degree (score) is performed, for example, as described in the above embodiment, between each reference model (including a silence model) constituting the word model and each frame feature parameter extracted from the audio signal. A matrix is constructed by mapping the likelihoods (see the upper part of FIGS. 1 to 3), and the highest score among the total scores of the likelihoods in various routes that go from the lower left corner to the upper right corner is determined as the matching degree. (Viterbi matching method).
[0068]
Here, as the likelihood of the silence model stored in the RAM, the highest likelihood of the likelihood group for the reference parameter of each frame feature parameter extracted in the frame period is defined as the likelihood of the silence model as in the above embodiment. , Or a method of replacing the average value of the top N items in each likelihood group with respect to the reference parameter of the frame feature parameter by the likelihood of the silent model. The likelihood calculated and replaced in this manner is the garbage likelihood calculated by the garbage likelihood calculation unit 10.
[0069]
Reference numeral 10 denotes a garbage likelihood calculating unit that calculates the likelihood for garbage by referring to the likelihood for each reference model stored in the RAM unit 6. Here, as the garbage likelihood, a method of setting the highest likelihood to the garbage likelihood among the likelihood group for the reference parameter of the frame feature parameter extracted in the frame period as in the above-described embodiment, Among the likelihoods of the frame feature parameter with respect to the reference parameter, a method of using the average value of the top N items as the likelihood for garbage is adopted.
[0070]
Reference numeral 11 denotes a recognition candidate storage unit, which compares the matching degrees (scores) calculated by the matching calculation unit 9 for each word, and stores M words having higher scores as recognition candidate words. Here, M words are recognized as recognition candidates for each dictionary (category) such as “size” and “information type”. The words of the recognition candidates stored in this way may be displayed as they are and the operator's intention may be selected, or, as described later, the M recognition candidates are again targeted. Alternatively, voice recognition processing may be performed.
[0071]
FIG. 7 is a block diagram showing the details of the matching calculation unit 9. Reference numeral 91 denotes a silence model likelihood determination unit which compares the likelihood of the silence model stored in the RAM unit 6 with the garbage likelihood from the garbage likelihood calculation unit 10 to determine the likelihood of the silence model.
[0072]
FIG. 8 shows a method of determining the likelihood of a silent model in the silent model determination unit 91. The likelihood of each reference model of the word model is determined in step S1 as to whether or not the reference model is a silent model. Here, when it is determined that the model is not a silent model, the likelihood of the model is stored in the RAM unit 6 as the likelihood of the word to be recognized.
[0073]
When it is determined in step S1 that the likelihood is the likelihood of the silence model, in steps S2 and S3, the likelihood of the silence model and the garbage likelihood from the garbage likelihood calculation unit 10 are calculated. It is determined which is larger, and if the garbage likelihood is larger, the likelihood of the silence model is replaced with the garbage likelihood in steps S5 and S6. For such replacement, the likelihood of the silence model in the RAM unit 6 may be rewritten to the likelihood for garbage, or the likelihood of the silence model in the RAM unit 6 may be rewritten without changing the likelihood. Only at the time of the calculation in step 9, the silence model may be processed to use the garbage likelihood.
[0074]
The speech recognition operation in the above embodiment will be described with reference to FIG.
[0075]
When the operator inputs a voice in the predetermined voice input mode, a dictionary to be used in the mode is selected from various dictionaries stored in the recognition dictionary unit 7, and one of the dictionaries is further selected. The dictionary is set as a recognition target dictionary (steps S101 and S102). When the dictionary to be recognized is set, one of the various words stored in the dictionary is read as the word to be recognized (W1) (steps S103 and S104). Then, this word (W1) is compared with the input speech signal as described above, and likelihood calculation and score calculation (matching processing) for word recognition are performed (step S105).
[0076]
When the score is calculated for the word (W1) read from the dictionary, the score is calculated as M words (Ws1, Ws2,..., Wsm) stored in the recognition candidate storage unit 11 by the previous processing. Is compared with the word with the lowest score. If the score is higher than this, the word (W1) is stored together with the score instead of the word stored earlier. Now, since the word (W1) is the first word read from the dictionary, the recognition candidate storage unit 11 has not yet stored the recognition candidate word. Therefore, the word (W1) is directly stored in the recognition candidate storage unit 11 together with the score (step S106).
[0077]
When the processing of the word (W1) is completed, the process returns to step S103, where the next word (W2) is read from the recognition dictionary, and the same processing as in steps S104 to S106 is performed. At this time, since M recognition candidates are not stored in the recognition candidate storage unit 11 until the M words are read from the dictionary, the words read from the dictionary are sequentially stored in the recognition candidate storage unit along with their scores. 11 is stored. When the word read from the dictionary reaches the (M + 1) th word, the score of this word (Wm + 1) is compared with the score of the M words stored in the recognition candidate storage unit 11, and if the score is larger than this. For example, this word (Wm + 1) and its score are stored in the recognition candidate storage unit 11, and the word with the lowest score and its score among the M words previously stored in the recognition candidate storage unit 11 are It is deleted from the recognition candidate storage unit 11.
[0078]
When the above processing is performed for all the words stored in the dictionary, it is determined in step S104 that the setting of the recognition candidates for the dictionary has been completed, and the processing returns to step S101. At this time, among the words stored in the dictionary, the top M words having a high score with respect to the speech input signal are stored in the recognition candidate storage unit 11 as recognition candidates.
[0079]
When the setting of the recognition candidates for the first dictionary is completed as described above, the next dictionary is selected as the dictionary to be recognized in steps S101 to S103, and the words in this dictionary are sequentially processed in steps S103 to S106. Is performed. Thereby, for the second dictionary, the top M words are stored in the recognition candidate storage unit 11 as recognition candidates.
[0080]
When the above operation is performed for all dictionaries to be used in the voice input mode, it is determined in step S102 that the speech recognition processing for all dictionaries has been completed. At this time, the recognition candidate storage unit 11 stores M words for each dictionary as recognition candidates for all dictionaries to be used in the voice input mode.
[0081]
Then, in step S107, the M recognition candidates are displayed for each dictionary section, for example, on a monitor of a voice recognition device. The operator selects a desired one from the recognition candidates displayed on the monitor. As a result, the word corresponding to the input voice is determined for each dictionary section.
[0082]
In the above speech recognition operation, M words are displayed on the monitor as recognition candidates for each dictionary section, and the operator is allowed to select one. However, when the number of words displayed as recognition candidates is large, the operator is forced to perform useless selection operations accordingly. It is preferable that the number of words to be displayed is as small as possible and that the accuracy of the word as a recognition candidate is high.
[0083]
Therefore, in the following embodiment, M words are not displayed as recognition candidates as they are, but the number of words is further reduced and the accuracy is increased as recognition candidates.
[0084]
FIG. 10 shows the configuration of this embodiment. The configuration of FIG. 10 is different from the embodiment of FIG. 4 only in the configuration of the word model creation unit 8 and the recognition candidate storage unit 11, and the other configuration is the same as the configuration of FIG.
[0085]
In the present embodiment, one word is selected from each dictionary section among the M words stored for each dictionary section in the recognition candidate storage unit 11 by the same processing as in the above-described embodiment, and this is selected as a silent model. The word model is created again by linking, and the matching between the word model and the input speech is calculated.
[0086]
FIG. 11 shows an example of a word model created by the word model creating section 8. This word model selects the word “gazou” from one dictionary section among the M words stored in the recognition candidate storage unit 11 for each dictionary section, and selects the word “large” from another dictionary section. They are selected and combined.
[0087]
For example, if there are two dictionaries to be used in accordance with the voice input mode, assuming that M words are set as recognition candidates for each dictionary in the same manner as in the processing of the above-described embodiment, one of the dictionary divisions The total number of word models selected and created one by one is M × M. Similarly, when there are three dictionaries to be used in accordance with the voice input mode, the total number of word models is M × M × M.
[0088]
In the present embodiment, likelihood calculation and matching processing with an input speech signal are performed for all of the word models of M raised to the Pth power (P is the number of dictionaries to be used according to the speech input mode). Is performed, L word models are discriminated from the highest score, and each word connected in this word model is set as a recognition candidate.
[0089]
When a word model is created by connecting a plurality of words in this way and compared with the input speech, whether one word or two words of each word model are included in the input speech, Depending on whether three words are included or not at all, that is, depending on the number of words included in the speech input signal, the difference in the matching score between each word model becomes large.
[0090]
Explaining this point in comparison with the above embodiment, in the above embodiment, a word model is created for only one word, and the degree of matching (score) between the word model and the input speech signal is calculated. . Therefore, the speech input signal always includes many unnecessary words in addition to the words of the word model. Therefore, the score of each word model is not so large even if the word is included in the input speech signal. Therefore, the difference in the degree of matching (score) between the word models does not increase so much. On the other hand, if a word model is created for a plurality of words as in the present embodiment and this is compared with the input speech signal to calculate the degree of matching (score), the word in the input speech Whether there is one or two words constituting the model, the difference in the scores between the word models is large. If all the words are completely included in the input speech signal, the score of the word model constituted by the words becomes extremely high.
[0091]
Therefore, when constructing a word model, it is better to construct a word model from a plurality of words as in the present embodiment than to construct a word model from one word as in the above embodiment. The difference between the scores of the candidates becomes large, and therefore, it becomes possible to provide the operator with highly accurate recognition candidate words.
[0092]
However, if a word model is created by connecting all the words one by one from all the word dictionaries used according to the input voice mode, the number of the word models becomes enormous. When matching processing with an input voice signal is performed on such a large number of word models, a huge processing time is required, and unnecessary processing on the word model due to unnecessary connection is repeated.
[0093]
Therefore, in the present embodiment, only the M words for each dictionary section obtained in FIGS. 4 to 9 are targeted, and one word is selected from each dictionary section and connected to create a word model. Then, by comparing this with the input voice signal, the number of final recognition candidates is reduced and the accuracy thereof is increased.
[0094]
Hereinafter, the operation of this embodiment will be described with reference to FIG. This operation is an operation in the case where there are two dictionaries to be used according to the voice input mode. In FIG. 12, the operations in steps S101 to S106 are the same as those in the above embodiment. That is, through these steps, M words are set as recognition candidates for each dictionary.
[0095]
When M words are set as recognition candidates for the two dictionaries to be used, the operation shifts from step S102 to step S201, and among these dictionaries, the recognition candidates set for the first dictionary are set. Is read out (steps S201 and S203), and the recognition candidate word (Ws21) set for the second dictionary is read out (steps S203 and S204). Then, these words (Ws11) and (Ws21) are connected by a silence model, and a silence model is further connected to both ends to create a word model (step S205).
[0096]
When the word model is created in this manner, likelihood calculation and score calculation (matching processing) with respect to the input speech signal are performed on this word model in the same manner as in the above embodiment (step S206). Then, this word model is stored in the recognition candidate storage unit 11 together with the score.
[0097]
When the processing for one word model is completed as described above, the process returns to step S203, and the word (W22) in the second dictionary is read. Then, this word (W22) is connected to the word (W11) in the first dictionary in the same manner as described above, and a new word model is created (step S205).
[0098]
Like the above, the generated word model is subjected to likelihood calculation and score calculation with the input speech signal (step S206), and is stored in the recognition candidate storage unit 11 together with the score.
[0099]
The operations in steps S203 to S206 described above are such that the M-th word (Ws2m) set for the second dictionary is linked to the word (Ws11) in the first dictionary to calculate a score, and this is stored in the recognition candidate storage unit 11. Repeated until memorized.
[0100]
When all of the M words set for the second dictionary are read out and the above processing ends, the process returns to step S201, where the next word (Ws12) set for the first dictionary is read out (step S201). S201, S202). Then, by repeating the processing of steps S203 to S206 in the same manner as described above, M words are sequentially connected to M words corresponding to the second dictionary, and M word models are created. The likelihood calculation and the score calculation between the input audio signals are sequentially performed. Then, the calculated scores are sequentially stored in the recognition candidate storage unit 11 together with the word model.
[0101]
The above process is repeated until all of the M words set for the first dictionary are connected to the M words of the second dictionary and processed.
[0102]
When the above processing is completed, the recognition candidate storage unit 11 stores a total of M × M word models and their scores. The scores of the M × M word models are compared in step S207, and the top L word models are selected. Then, words in each dictionary included in the upper L word models are determined, and the words are displayed on a monitor as recognition candidates for each dictionary.
[0103]
In this embodiment, the operation is performed when two dictionaries are used according to the voice input mode. However, the operation is not limited to this. For example, when there are three dictionaries, steps corresponding to steps S201 and S202 (for the first dictionary) and steps S203 and S204 (for the second dictionary) in FIG. You just need to add a step. As the number of target dictionaries increases, such a step may be added so that all M words corresponding to each dictionary can be combined.
[0104]
Even when there are three or more target dictionaries (for example, K), instead of selecting words one by one from the K dictionaries, J dictionaries (J <K) are selected. May be selected, and words corresponding to the selected J dictionaries may be selected one by one and connected.
[0105]
Further, in this embodiment, M word models set for each dictionary are combined to create a word model of M to the Pth power (P is the number of dictionaries). Null (none) may be added as a word in addition to the M words, and M + 1 words set for each dictionary may be used to create M + 1 P-th word models. In this case, a combination of a null and a word is performed by connecting words except for the null. For example, if there are three target dictionaries, a word in the first dictionary is null, a word in the second dictionary is Ws1, and a word in the third dictionary is Ws2, a word model combining these is: The word Ws1 and the word Ws2 are connected by a silence model, and a silence model is further connected to both ends thereof. When there are two target dictionaries, the first dictionary is null, and the second dictionary is Ws2, the word model is a word model in which a silence model is connected to both ends of the word Ws2. It becomes.
[0106]
As described above, if nulls are separately added in addition to the M words, even if the operator does not input words for all of the types and divisions that are required to be input by the voice input mode, words of the input type are not input. Can be recognized correctly. For example, it is assumed that when the voice input mode requires input of words of three types / sections of A, B, and C, the operator has input only words of the types / sections of A and B. In this case, in the embodiment of FIG. 12, in steps S101 to S106, M words are set as recognition candidates for dictionaries corresponding to the types and divisions of A, B, and C. Of these, the dictionary for C is set. Since the M words that have been selected correspond to the types and distinctions that have not been input by the operator, all of them are erroneous as recognition candidates. However, in the embodiment of FIG. 12, the words of the recognition candidates are set for the dictionary of C by steps S201 to S206, and are displayed on the monitor.
[0107]
Therefore, if nulls are further added to the M words set for the dictionaries A, B, and C, the score of the word model when null is selected for the dictionary of C becomes higher than the others. In other words, the word model in this case is O + Wa + O + Wb + O (O is a silent model), where the words in the A and B dictionaries are Wa and Wb, respectively. Since it is in accordance with the classification, the audio part of Wa and A and the audio part of Wb and B are matched, and the overall score is increased.
[0108]
In the case of a matching method in which the score is proportional to the length of the word model, if the null is selected, the length of the word model is reduced, so that the score needs to be normalized. Such normalization is achieved, for example, by averaging scores according to the length of the word model.
[0109]
This point is the same even when only one word is targeted as in the embodiment of FIG. That is, the number of syllables of a word is not uniform, and the number of syllables differs depending on the word. For example, “Gamen” has three syllables, and “Onsei” has four syllables. In such a case as well, the length of the word model changes according to the number of syllables, but the scores are averaged according to the length of the word model by the normalization process. The gap is corrected.
[0110]
Although various embodiments according to the present invention have been described above, the present invention is not limited to such embodiments.
[0111]
For example, in the above embodiment, when a word model is created from one word, only one silence model is added to both ends of the word. However, two or more silence models may be added. The number of silence models may be changed before and after.
[0112]
In addition, when a word model is created by connecting two or more words, the number of silence models interposed between words is one in the above embodiment, but this can be two or more. Words may be directly linked without a silence model. Further, the number of silence models interposed between the words Wa and Wb is two, the number of silence models intervening between the words Wb and Wc is one, and so on. May be changed.
[0113]
In the above-described embodiment, the largest likelihood or the average value of the top N likelihoods among the likelihoods of the frame feature amount with respect to the reference model feature amount is adopted as the garbage likelihood. The K-th likelihood may be set as the garbage likelihood. At this time, if K is selected so that the K-th likelihood is statistically close to the average value of the N pieces of likelihood, the same effect as when the average value is employed without averaging is obtained. Can be
[0114]
Further, in the above-described embodiment, the silence model is added by the word model creation unit 8. Alternatively, a silence model may be added to the word in advance and stored in the recognition dictionary unit 7. .
[0115]
Further, in the above embodiment, in addition to the M words set as recognition candidates for each dictionary, nulls are separately added to connect each word. In this case, nulls are set for all dictionaries. Then, the word model consists of only the silence model. Therefore, all null word models may be excluded from matching. Or. If all of them are null and the matching score is higher than the top H, the processing result for the input voice may not be adopted and the operator may be prompted to input voice again.
[0116]
In the above embodiment, the number of recognition candidates set for each dictionary is uniformly set to M, but the number of recognition candidates may be changed for each dictionary. At this time, the number of recognition candidates may be set in advance for each dictionary, or the number of recognition candidates for the dictionary may be set according to the score at the time of recognition processing. In the latter case, for example, a threshold value of the score may be set, and only those having a score equal to or larger than the threshold value may be set as recognition candidates. In this case, the number of recognition candidates depends on the score and the threshold, and may be more than or less than M.
[0117]
In the above embodiment, for example, in FIG. 12, the characteristic analysis and calculation processing in step S105 and the characteristic analysis and calculation processing in step S206 are the same, but the characteristic analysis and calculation processing in step S105 are roughened. The characteristic analysis and calculation processing in step S206 may be made precise. That is, in step S105, the number of target word models is large, so that the processing speed is prioritized by coarse processing. In step S206, the number of target word models is small, and the accuracy is increased by dense processing. This makes it possible to obtain an accurate recognition result while increasing the overall processing speed.
[0118]
Here, the recognition processing accuracy distinguishes between coarse processing and dense processing by changing parameters to be processed with respect to acoustic analysis parameters such as a spectrum of an audio signal, a change in spectrum, power and a change in power. . For example, the coarse processing targets only the parameters of the spectrum, and the dense processing targets the spectrum, the amount of change in the spectrum, the power, and the amount of change in the power. Alternatively, the number of extracted frames of the input audio signal may be changed between the coarse processing and the fine processing. For example, when the number of frames for dense processing is set to 100, the number of frames for coarse processing is reduced to 50.
[0119]
In addition, various changes can be made to the characteristic analysis and the matching process. In addition, the garbage model can be generated in various ways other than the method of obtaining the maximum likelihood of the frame and the method of obtaining the average of the upper N items as described above.
[0120]
【The invention's effect】
According to the present invention, the likelihood of the silence model for the frame feature is appropriately replaced with the likelihood for garbage. Therefore, even if the input speech signal including the silence portion is cut out, the matching calculation for the silence portion is not performed. The effect is absorbed by the silence model, and the effect on the matching operation of the unnecessary word portion in the audio signal is absorbed by the replacement with the garbage likelihood. , Can perform voice recognition.
[0121]
Further, since the garbage likelihood is calculated based on the likelihood of the frame feature amount with respect to the reference model feature amount, even if the word to be recognized is added to the dictionary, all the likelihoods are reduced as in the conventional example. It is not necessary to separately calculate and reset based on the word, and the degree of freedom of the device in adding or changing a word can be improved.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an outline of an embodiment;
FIG. 2 is a diagram illustrating an outline of an embodiment;
FIG. 3 is a diagram for explaining an outline of an embodiment;
FIG. 4 is a diagram showing a configuration of an embodiment.
FIG. 5 is a diagram illustrating a storage state of a recognition dictionary unit according to the embodiment;
FIG. 6 is a diagram showing a configuration of a word model according to the embodiment.
FIG. 7 is a diagram illustrating a configuration of a matching calculation unit according to the embodiment;
FIG. 8 is a diagram illustrating a method of setting a likelihood of a silence model according to the embodiment;
FIG. 9 is a diagram showing the operation of the embodiment.
FIG. 10 is a diagram showing a configuration of a second embodiment.
FIG. 11 is a diagram showing a configuration of a word model according to a second embodiment.
FIG. 12 is a diagram showing the operation of the second embodiment.
[Description of sign]
1 Voice input section
2 Audio signal extraction unit
3 Sound analysis unit
4 Reference model parameter section
5 Likelihood calculation unit
6 RAM section
7 Recognition dictionary
8 Word model creation unit
9 Matching operation unit
10 Garbage likelihood calculator
11 Recognition candidate storage unit

Claims (10)

Acoustic analysis means for acoustically analyzing an input voice signal to extract frame features; word model creation means for creating a word model by connecting silence models to both ends of a word reference model; A likelihood calculating means for calculating the likelihood of the frame feature with respect to the word model by comparing with the feature of the word model, and calculating a matching degree of the word model with respect to the input speech signal based on the calculated likelihood. Matching calculation means, a recognition candidate setting means for setting a recognition candidate according to the degree of matching, and a garbage likelihood setting means for setting a garbage likelihood for the silent model,
The garbage likelihood setting means calculates the garbage likelihood by averaging the likelihoods of the frame features from the top model to the Nth among the likelihoods with respect to the reference model feature,
The speech recognition device according to claim 1, wherein said matching calculation means performs a matching calculation using the garbage likelihood set by said garbage likelihood setting means as a likelihood of a silence model. Acoustic analysis means for acoustically analyzing an input voice signal to extract frame features; word model creation means for creating a word model by connecting silence models to both ends of a word reference model; A likelihood calculating means for calculating the likelihood of the frame feature with respect to the word model by comparing with the feature of the word model, and calculating a matching degree of the word model with respect to the input speech signal based on the calculated likelihood. Matching calculation means, a recognition candidate setting means for setting a recognition candidate according to the degree of matching, and a garbage likelihood setting means for setting a garbage likelihood for the silent model,
The garbage likelihood setting means calculates the garbage likelihood by averaging the likelihoods of the frame features from the top model to the Nth among the likelihoods with respect to the reference model feature,
The matching calculating means selects one of the likelihood of the silence model calculated by the likelihood calculating means and the garbage likelihood set by the garbage likelihood setting means as the likelihood of the silence model. A speech recognition device that performs a matching operation. Acoustic analysis means for acoustically analyzing an input voice signal to extract frame features; word model creation means for creating a word model by connecting silence models to both ends of a word reference model; A likelihood calculating means for calculating the likelihood of the frame feature with respect to the word model by comparing with the feature of the word model, and calculating a matching degree of the word model with respect to the input speech signal based on the calculated likelihood. Matching calculation means, a recognition candidate setting means for setting a recognition candidate according to the degree of matching, and a garbage likelihood setting means for setting a garbage likelihood for the silent model,
The garbage likelihood setting means, among the likelihoods of the frame feature amount with respect to the reference model feature amount, sets the Kth likelihood from the top in the magnitude of the likelihood as the garbage likelihood,
The speech recognition device according to claim 1, wherein said matching calculation means performs a matching calculation using the garbage likelihood set by said garbage likelihood setting means as a likelihood of a silence model. Acoustic analysis means for acoustically analyzing an input voice signal to extract frame features; word model creation means for creating a word model by connecting silence models to both ends of a word reference model; A likelihood calculating means for calculating the likelihood of the frame feature with respect to the word model by comparing with the feature of the word model, and calculating a matching degree of the word model with respect to the input speech signal based on the calculated likelihood. Matching calculation means, a recognition candidate setting means for setting a recognition candidate according to the degree of matching, and a garbage likelihood setting means for setting a garbage likelihood for the silent model,
The garbage likelihood setting means, among the likelihoods of the frame feature amount with respect to the reference model feature amount, sets the Kth likelihood from the top in the magnitude of the likelihood as the garbage likelihood,
The matching calculation means selects one of the likelihood of the silence model calculated by the likelihood calculation means and the garbage likelihood set by the garbage likelihood setting means as the likelihood of the silence model. A speech recognition device that performs a matching operation. The matching calculation means according to claim 2 or 4, wherein the matching calculation means comprises: A speech recognition apparatus characterized in that the likelihood of a silence model calculated by a stage is compared with the garbage likelihood set by the garbage likelihood setting means, and the larger likelihood is selected. Extracting a frame feature by acoustically analyzing the input voice signal; connecting a silence model to both ends of the reference model of the word to create a word model; and determining the frame feature and the feature of the reference model. Calculating the likelihood of the frame feature amount for the word model by comparing; calculating the matching degree of the word model with respect to the input speech signal based on the calculated likelihood; Setting a recognition candidate accordingly, and setting a garbage likelihood for the silence model,
The garbage likelihood setting step calculates the garbage likelihood by averaging the likelihoods of the likelihood from the top to the Nth among the likelihoods of the frame feature amount with respect to the reference model feature amount,
The voice recognition method according to claim 1, wherein the matching calculation step performs the matching calculation using the garbage likelihood set in the garbage likelihood setting step as the likelihood of the silence model. Extracting a frame feature by acoustically analyzing the input voice signal; connecting a silence model to both ends of the reference model of the word to create a word model; and determining the frame feature and the feature of the reference model. Calculating the likelihood of the frame feature amount for the word model by comparing; calculating the matching degree of the word model with respect to the input speech signal based on the calculated likelihood; Setting a recognition candidate accordingly, and setting a garbage likelihood for the silence model,
The garbage likelihood setting step calculates a garbage likelihood by averaging the likelihood of the likelihood from the top to the N-th among the likelihoods of the frame feature amount with respect to the reference model feature amount,
The matching calculation step selects one of the likelihood of the silence model calculated by the likelihood calculation step and the garbage likelihood set by the garbage likelihood setting step as the likelihood of the silence model. A speech recognition method characterized by performing a matching operation by using Extracting a frame feature by acoustically analyzing the input voice signal; connecting a silence model to both ends of the reference model of the word to create a word model; and determining the frame feature and the feature of the reference model. Calculating the likelihood of the frame feature amount for the word model by comparing; calculating the matching degree of the word model with respect to the input speech signal based on the calculated likelihood; Setting a recognition candidate accordingly, and setting a garbage likelihood for the silence model,
In the garbage likelihood setting step, among the likelihoods of the frame feature amount with respect to the reference model feature amount, the K-th likelihood having the highest likelihood as the garbage likelihood is defined as garbage likelihood,
The voice recognition method according to claim 1, wherein the matching calculation step performs the matching calculation using the garbage likelihood set in the garbage likelihood setting step as the likelihood of the silence model. Extracting a frame feature by acoustically analyzing the input voice signal; connecting a silence model to both ends of the reference model of the word to create a word model; and determining the frame feature and the feature of the reference model. Calculating the likelihood of the frame feature amount for the word model by comparing; calculating the matching degree of the word model with respect to the input speech signal based on the calculated likelihood; Setting a recognition candidate accordingly, and setting a garbage likelihood for the silence model,
In the garbage likelihood setting step, among the likelihoods of the frame feature amount with respect to the reference model feature amount , the K-th likelihood having the highest likelihood as the garbage likelihood is defined as garbage likelihood,
In the matching calculation step, as the likelihood of the silence model, one of the likelihood of the silence model calculated in the likelihood calculation step and the garbage likelihood set in the garbage likelihood setting step is selected. A speech recognition method characterized by performing a matching operation by using 10. The matching calculation step according to claim 7 , wherein the likelihood of the silence model calculated in the likelihood calculation step is compared with the garbage likelihood set in the garbage likelihood setting step . A speech recognition method characterized by selecting likelihood.

Images