JP4678464B2 - Voice recognition apparatus, voice recognition method, program, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform a voice recognition processing at a speed or with a precision meeting the requirement of a user or the like. SOLUTION: A partial space detection part 21 detects partial space which a feature vector of a voice belongs to in its feature vector space. A calculation object function selection part 23 stores a calculation object function table where one or more probability density functions which define HMM for use in matching processing with feature vectors are made correspond to each of a lot of partial spaces of the feature vector space, and the calculation object function selection part 23 selects one corresponding to speed/prevision information from a speed/precision setting part 27 from probability density functions corresponding to the partial space outputted from the partial space detection part 21. A score calculation part 25 uses the selected probability density function to perform matching processing between the feature vector and HMM.

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a voice recognition device, a voice recognition method, a program, and a recording medium, and, for example, a voice recognition device and a voice recognition method capable of performing voice recognition processing with speed or accuracy according to a request from a user or the like. And a program and a recording medium.
[0002]
[Prior art]
FIG. 1 shows an example of the configuration of a conventional speech recognition apparatus.
[0003]
The voice uttered by the user is input to a microphone (microphone) 1, and the microphone 1 converts the input voice into a voice signal as an electrical signal. This audio signal is supplied to an AD (Analog Digital) converter 2. In the AD conversion unit 2, the audio signal that is an analog signal from the microphone 1 is sampled, quantized, and converted into audio data that is a digital signal. This audio data is supplied to the feature extraction unit 3.
[0004]
The feature extraction unit 3 performs acoustic processing on the audio data from the AD conversion unit 2 for each appropriate frame, thereby extracting, for example, a feature vector (feature amount) such as MFCC (Mel Frequency Cepstrum Coefficient), This is supplied to the matching unit 4. In addition, the feature extraction unit 3 can extract other feature quantities such as a spectrum, a linear prediction coefficient, a cepstrum coefficient, and a line spectrum pair.
[0005]
The matching unit 4 uses the feature vector from the feature extraction unit 3 to refer to the acoustic model database 5, the dictionary database 6, and the grammar database 7 as necessary, and the voice (input voice) input to the microphone 1. Are recognized based on, for example, a continuous distribution HMM method.
[0006]
That is, the acoustic model database 5 stores an acoustic model representing acoustic features such as individual phonemes and syllables in the speech language for speech recognition. Here, since speech recognition is performed based on the continuous distribution HMM method, for example, an HMM (Hidden Markov Model) is used as the acoustic model. The dictionary database 6 stores a word dictionary in which information related to pronunciation (phonological information) is described for each word to be recognized. The grammar database 7 stores grammatical rules (language model) describing how each word registered in the word dictionary of the dictionary database 6 is linked (connected). Here, as the grammar rule, for example, a rule based on context-free grammar (CFG), statistical word chain probability (N-gram), or the like can be used.
[0007]
The matching unit 4 connects the acoustic model stored in the acoustic model database 5 by referring to the word dictionary in the dictionary database 6 to construct an acoustic model (word model) of the word. Further, the matching unit 4 connects several word models by referring to the grammar rules stored in the grammar database 7, and the word model thus connected and the voice input to the microphone 1. A matching process with the extracted feature vector series is performed based on, for example, a continuous distribution HMM method, and the speech is recognized. That is, the matching unit 4 detects the sequence of the word model having the highest score (likelihood) from which the time-series feature vector supplied from the feature extraction unit 3 is output (observed), and corresponds to the sequence of the word model. The word string to be output is output as a speech recognition result.
[0008]
That is, the matching unit 4 accumulates the appearance probabilities of the feature vectors for the word strings corresponding to the connected word models, uses the accumulated value as a score, and sets the word string having the highest score as the speech recognition result. Output.
[0009]
The score calculation is generally performed by an acoustic score given by an acoustic model stored in the acoustic model database 5 (hereinafter, referred to as an acoustic score as appropriate) and a linguistic score given by a grammar rule stored in the grammar database 7. (Hereinafter referred to as language score as appropriate).
[0010]
That is, as the acoustic score, for example, in the case of the HMM method, the cumulative probability of outputting (observing) the feature vector series output by the feature extracting unit 3 from the acoustic model constituting the word model. The value is calculated. As the language score, for example, in the case of bigram, the probability that the word of interest and the word immediately preceding the word are linked (connected) is obtained. Then, a speech recognition result is determined based on a final score (hereinafter, referred to as final score as appropriate) obtained by comprehensive evaluation (for example, weighted addition) of the acoustic score and language score for each word. .
[0011]
By performing the processing as described above, in the speech recognition apparatus of FIG. 1, for example, when the user speaks “Today is a good weather”, “Today”, “Ha”, “Good”. , “Weather”, “Issue” are given an acoustic score and a language score, and when the final score obtained by comprehensive evaluation is the largest, the word strings “today”, “ha”, “ “Good”, “Weather”, and “It is” are output as the speech recognition results.
[0012]
By the way, in the above-mentioned case, if five words “today”, “ha”, “good”, “weather”, and “sound” are registered in the word dictionary of the dictionary database 6, these five words are registered. The sequence of 5 words that can be constructed using words is 5^FiveExist. Therefore, simply, in the matching unit 4, this 5^FiveIt is necessary to evaluate the street word strings and determine the one that best matches the user's utterance (the one that maximizes the final score). If the number of words to be registered in the word dictionary increases, the number of words arranged as many as the number of words becomes the number of words multiplied by the number of words. Therefore, the number of word strings to be evaluated is enormous. It becomes.
[0013]
Furthermore, in general, since the number of words included in an utterance is unknown, not only a word string consisting of a sequence of five words but also a word string consisting of one word, two words,. There is a need to. Therefore, since the number of word strings to be evaluated is further enormous, the most probable speech recognition result is efficiently determined from the viewpoint of the amount of computation out of such enormous word strings. That is a very important issue.
[0014]
For example, E. Bocchieri.Vector quantization for the efficient computation of continuous density likelihoods.In International Conference on Acoustic, Speech, and Signal Processing, volume 2, pages 692-695, Apr. 1993 (hereinafter referred to as Reference 1), KMKnill, MJFGales, and SJYoung. Use of gaussian selection in large vocabulary continuous speech recognition using hmms. In International Conference on Spoken Language Processing, volume 1, pages 470-473, Oct. 1996 (hereinafter referred to as Reference 2), MJFGales, KMKnill, and SJYoung. State-based gaussian selection in large vocabulary continuous speech recognition using hmms. In Cambridge University Technical Report, TR284, Jan. 1997 (hereinafter referred to as Reference 3), SMHerman and RASukkar. Variable threshold vector quantization for reduced continuous density likelihood computation in speech recognition. In IEEE Workshop on Acou As described in stic Speech Recognition and Understanding Proceedings, pages 331-338, Santa Babara, 1997 (hereinafter referred to as Document 4), there is a method of omitting a part of the calculation of the output probability constituting the acoustic score.
[0015]
That is, for example, according to the continuous HMM method, the HMM has a feature vector x at time t in a state s._tOutput probability b_s(X_t) Is calculated by the following equation.
[0016]
b_s(X_t) = Σc_m× g_m(X_t) ... (1)
[0017]
Here, in equation (1), c_mIs the mth function g_mIs the weighting factor for () and the function g_m() Is an m-th probability density function (for example, Gaussian distribution) constituting the HMM. Σ represents a summation for the variable m. Therefore, according to equation (1), the output probability b_s(X_t) Is the probability density function g_m(X_t) Weight c_mCalculated as a sum.
[0018]
Weight coefficient c_mAnd probability density function g_m() Is one piece of definition information as information that defines the HMM that is an acoustic model (for example, other definition information includes a state of the HMM from a certain state and other information including the state as necessary. State transition probability as the probability of transition to a state), the HMM is a weighting factor c_mAnd probability density function g_mIn addition to the case where the set of () is defined using only one set, there are cases where it is defined using a plurality of sets.
[0019]
The HMM has a plurality of N sets of weighting factors c₀Thru c_N-1And probability density function g₀To g_N-1When defined using (), the calculation of Equation (1) needs to be performed by changing the variable m to an integer value from 0 to N-1.
[0020]
However, the N probability density functions g that define the HMM₀(X_t), G₁(X_t), ..., g_N-1(X_t) Includes output probability b_s(X_t) As a result, there may be a case where the magnitude (degree) contributing to the acoustic score (hereinafter, referred to as the degree of contribution as appropriate) is very small (that contributes little).
[0021]
Therefore, in the methods described in Documents 1 to 4, the probability density function g with a very small contribution is used._mBy omitting () from the calculation of equation (1), the amount of calculation is reduced while suppressing deterioration of speech recognition accuracy.
[0022]
Specifically, for example, for each predetermined partial space of the feature vector space as shown in FIG._tOutput probability b_s(X_t1 or more probability density function g used in the calculation of_mA table in which () is associated (hereinafter referred to as a calculation target function table as appropriate) is created, and a feature vector x belonging to a certain partial space_tOutput probability b_s(X_t) Is calculated by N probability density functions g₀() To g_N-1In (), in the calculation target function table, the feature vector x_tThis is performed using only those associated with the subspace to which the belongs.
[0023]
In this case, calculation of a part of the probability density functions can be omitted, so that the amount of calculation is reduced and the speed of the speech recognition processing can be improved. In addition, the probability density function for which the calculation is omitted is the score b_s(X_t), The degradation of speech recognition accuracy due to omission of the calculation can be suppressed.
[0024]
Here, in the calculation target function table of FIG. 2 (the same applies to FIGS. 6 to 9 described later), the feature vector space is Y₀, Y₁, ..., Y₅₁₁Are divided into 512 subspaces.
[0025]
In the calculation target function table, not the feature space x but the feature vector x_tFor each feature vector x_tOutput probability b_s(X_t1 or more probability density function g used in the calculation of_mIt is ideal to associate (), but then the feature vector x_tSince the calculation target function table cannot be created because is a continuous quantity, the calculation target function table has a probability density function g for each subspace._mIt is created in a form that associates.
[0026]
According to the calculation target function table of FIG.₀Feature vector x belonging to_tOutput probability b_s(X_t) Is the formula b_s(X_t) = C₀g₀(X_t) + C₁g₁(X_t) + C_Fiveg_Five(X_t) + C₁₅g₁₅(X_t).
[0027]
On the other hand, the probability density function g that defines the HMM_mAssuming that the total number N of () is 16, for example, when the formula (1) is adopted as it is, the feature vector x_tOutput probability b_s(X_t) Is the formula b_s(X_t) = C₀g₀(X_t) + C₁g₁(X_t) + ... + c₁₅g₁₅(X_t).
[0028]
Therefore, when using the formula (1) as it is, it is necessary to perform calculations for 16 probability density functions. However, when using the calculation target function table, it is sufficient to perform calculations for four probability density functions. Therefore, the amount of calculation can be greatly reduced.
[0029]
When using the calculation target function table, the feature vector x_tIs 512 subspace Y₀To Y₅₁₁It is necessary to detect which one of the subspaces belongs, but as a method for detecting this subspace, for example, vector quantization can be used.
[0030]
[Problems to be solved by the invention]
As described above, by using the calculation target function table, it is possible to reduce the amount of calculation for obtaining the acoustic score (output probability), and to improve the speech recognition processing speed while suppressing a decrease in speech recognition accuracy.
[0031]
However, for example, even when the resources allocated to the voice recognition process are reduced, when the voice recognition process in real time is required, even if there is a slight deterioration in the voice recognition accuracy, the voice recognition processing speed In other words, it is desirable to reduce the amount of calculation according to the resource.
[0032]
This is because when real-time speech recognition processing is required, processing is generally performed based on the speech recognition result, and therefore, when the speech recognition result cannot be obtained in real time. This is because it will hinder subsequent processing.
[0033]
On the other hand, for example, when there are sufficient resources that can be allocated to the speech recognition processing, it is possible to obtain a highly accurate speech recognition processing in real time by performing many calculations with the sufficient resources. That is, in this case, even if the speech recognition processing speed is reduced, a highly accurate speech recognition result can be obtained in real time. Therefore, when resources are sufficient, it is desirable from the viewpoint of effective use of resources to perform highly accurate speech recognition processing using the sufficient source.
[0034]
The present invention has been made in view of such a situation, and makes it possible to perform speech recognition processing with speed and accuracy according to a request.
[0035]
[Means for Solving the Problems]
  A speech recognition device, program, or recording medium according to an aspect of the present invention is a speech recognition device that recognizes speech, and the extraction feature that extracts the feature amount of the speech, and the feature amount of the speech includes the feature. Detection means for detecting a subspace belonging to the quantity space, and one or more HMMs (Hidden Markov Models) used for matching processing with the feature quantity of the speech for each of the plurality of subspaces of the feature quantity space Selection means for selecting any one or more definition information from storage means for storing definition information in association with the one or more definition information associated with the partial space to which the audio feature value belongs Then, using the definition information selected by the selection means, the voice corresponds to the HMM by performing a matching process between the feature quantity of the voice and the HMM. And a matching unit that outputs a speech recognition result of the speech based on the score, and the definition information obtains an output probability that the HMM outputs the feature quantity. A probability density function or a probability function used in the above, wherein the selection means is set according to the speed or accuracy of the voice recognition process set according to the user's operation, or a resource that can be allocated to the voice recognition process A speech recognition device that selects a number of the definition information corresponding to the speed or accuracy of the speech recognition processing in an order corresponding to the magnitude that the definition information contributes to the score, based on the speed or accuracy of the recognition processing; As such a voice recognition device, a program for causing a computer to function, or a recording medium on which such a program is recorded That.
[0036]
  A speech recognition method according to an aspect of the present invention is a speech recognition method of a speech recognition device that recognizes speech, wherein the speech recognition device extracts the feature amount of the speech, and the feature amount of the speech is A detection step of detecting a subspace belonging to the feature amount space, and an HMM (Hidden Markov Model) used for matching processing with the speech feature amount for each of the plurality of subspaces of the feature amount space Any one or more definition information is selected from the one or more definition information associated with the partial space to which the audio feature amount belongs in a storage unit that stores one or more definition information in association with each other. And performing a matching process between the audio feature quantity and the HMM using the definition information selected in the selection step, and A matching step for obtaining a score representing the likelihood of being compatible with the HMM and outputting a speech recognition result of the speech based on the score, and the definition information includes the feature amount by the HMM. Probability density function or probability function used to determine the output probability to be output, and in the selection step, the speed or accuracy of speech recognition processing set according to the user's operation, or resources that can be allocated to speech recognition processing The number of the definition information corresponding to the speed or accuracy of the voice recognition processing based on the speed or accuracy of the voice recognition processing set according to The voice recognition method to select.
[0039]
  In one aspect of the present invention, the feature amount of the voice is extracted, and a partial space to which the feature amount of the voice belongs in the feature amount space is detected. Further, each of the plurality of partial spaces of the feature amount space is stored in association with one or more definition information defining an HMM (Hidden Markov Model) used for the matching processing with the feature amount of the speech. Any one or more pieces of definition information are selected from the one or more pieces of definition information associated with the partial space to which the audio feature amount belongs. Then, by using the selected definition information to perform a matching process between the voice feature quantity and the HMM, a score representing the likelihood that the voice corresponds to the HMM is obtained. And the speech recognition result of the speech is output based on the score. The definition information is a probability density function or a probability function that is used to obtain an output probability that the HMM outputs the feature value. In the selection of the definition information, a speech recognition process set according to a user operation is performed. Based on speed or accuracy, or speed or accuracy of speech recognition processing set according to resources that can be allocated to speech recognition processing, the number of the definition information corresponding to the speed or accuracy of the speech recognition processing, The definition information is selected in the order corresponding to the size that contributes to the score.
[0040]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 3 shows a configuration example of an embodiment of a speech recognition apparatus to which the present invention is applied. In the figure, portions corresponding to those in FIG. 1 are denoted by the same reference numerals, and description thereof will be omitted below as appropriate. That is, the speech recognition apparatus in FIG. 3 is configured in the same manner as in FIG. 1 except that a matching unit 11 is newly provided instead of the matching unit 4.
[0041]
FIG. 4 shows a configuration example of the matching unit 11 of FIG.
[0042]
The time-series feature vector output from the feature extraction unit 3 (FIG. 3) is supplied to the subspace detection unit 21 and the score calculation unit 25.
[0043]
The subspace detection unit 21 refers to the subspace data storage unit 22, detects the subspace to which the feature vector supplied thereto belongs in the feature vector space, and displays subspace information representing the subspace, The calculation target function selection unit 23 is supplied.
[0044]
The subspace data storage unit 22 stores subspace data as information necessary for the subspace detection unit 21 to detect the subspace to which the feature vector belongs.
[0045]
Here, the subspace detection unit 21 can detect the subspace to which the feature vector belongs by, for example, vector quantization. In this case, the subspace data storage unit 22 uses the subspace data as the subspace data. A code book used for the vector quantization is stored.
[0046]
The code book can be created by using a large number of audio data and performing learning using, for example, an LBG (Linde Buzo Gray) algorithm, which is one of code book learning algorithms.
[0047]
The code book includes a code as a representative vector of each subspace when the feature vector space is divided into several subspaces (in this embodiment, as described above, 512 subspaces). A vector and a code representing the code vector are registered. Therefore, when the feature vector space is divided into, for example, 512 partial spaces, codes corresponding to the 512 code vectors are registered in the code book.
[0048]
The subspace detection unit 21 calculates the distance between the feature vector and each of the 512 code vectors registered in the codebook, and detects the code vector that minimizes the distance. Then, the subspace detection unit 21 assumes that the subspace whose representative vector is the code vector is the subspace to which the feature vector belongs, and converts the code corresponding to the detected code vector to the subspace to which the feature vector belongs. Output as subspace information.
[0049]
The calculation target function selection unit 23 refers to the calculation target function table (definition information table) stored in the calculation target function table storage unit 24 based on the subspace information from the subspace detection unit 21, thereby determining the feature vector. A probability density function or the like that defines an acoustic model (here, as described above, HMM) to be used for calculation of the used acoustic score (output probability) is selected.
[0050]
That is, in addition to the partial space information supplied from the partial space detection unit 21, the calculation target function selection unit 23 is also supplied with speed / accuracy information from the speed / accuracy setting unit 27.
[0051]
The calculation target function selection unit 23 uses a probability density function or the like associated with the subspace represented by the subspace information from the subspace detection unit 21 in the calculation target function table stored in the calculation target function table storage unit 24. Based on the speed / accuracy setting information supplied from the speed / accuracy setting unit 27, one or more probability density functions and the like are selected. Then, the calculation target function selection unit 23 supplies selection information representing the selected probability density function or the like to the score calculation unit 25.
[0052]
The calculation target function table storage unit 24 generates a calculation target function table in which one or more probability density functions that define the acoustic model stored in the acoustic model database 5 are associated with each of the plurality of partial spaces of the feature vector space. I remember it.
[0053]
  The score calculation unit 25 uses the feature vector supplied from the feature extraction unit 3 to store the acoustic model stored in the acoustic model database 5, the word dictionary stored in the dictionary database 6, and the grammar stored in the grammar database 7. While referring to the rules as necessary, constructing speech recognition result candidates (hereinafter referred to as hypotheses as appropriate)hypothesis, An acoustic score based on the HMM method as described above and a language score are calculated.
[0054]
However, the score calculation unit 25 uses the probability density function represented by the selection information supplied from the calculation target function selection unit 23 instead of all the probability density functions that define the acoustic model for the acoustic score._tOutput probability b_s(X_t) And an acoustic score is obtained based on the output probability.
[0055]
The acoustic score and language score obtained in the score calculation unit 25 are supplied to the output selection unit 26, and the output selection unit 26 comprehensively evaluates the acoustic score and language score obtained for each hypothesis to obtain a final score. For example, a hypothesis that maximizes the final score is selected and output as a speech recognition result.
[0056]
The speed / accuracy setting unit 27 sets the speed or accuracy of the voice recognition processing according to the operation of the operation lever 28 and supplies speed / accuracy information representing the set speed or accuracy to the calculation target function selection unit 23. .
[0057]
The operation lever 28 is operated when the user specifies the speed or accuracy of the voice recognition processing, and supplies an operation signal corresponding to the operation to the speed / accuracy setting unit 27.
[0058]
Accordingly, the speed / accuracy setting unit 27 sets the speed or accuracy of the speech recognition process in accordance with the user's request.
[0059]
Here, the operation lever 28 can be configured as a physical lever, or can be configured as a virtual lever displayed on the screen. In the case where the operation lever 28 is configured as a physical lever, the operation lever 28 is actually gripped and operated by the user. When the operation lever 28 is configured as a virtual lever, the operation lever 28 is operated by the user by dragging with the mouse.
[0060]
In the embodiment shown in FIG. 4, when a lower speed or higher accuracy voice recognition process is requested, the operation lever 28 is operated in the left direction, and conversely, a higher speed or lower accuracy voice recognition process is performed. When requested, the operation lever 28 is operated in the right direction.
[0061]
Next, the matching process performed by the matching unit 11 in FIG. 4 will be described with reference to the flowchart in FIG.
[0062]
When the user speaks and the feature extraction unit 3 starts outputting the feature vector of the voice, the matching unit 11 starts the matching process.
[0063]
That is, the time-series feature vector output from the feature extraction unit 3 is supplied to the subspace selection unit 21 and the score calculation unit 25, and the subspace detection unit 21 refers to the subspace data storage unit 22 in step S1. , Feature vector x from feature extraction unit 3_tDetect the subspace to which the belongs. Then, the partial space detection unit 21 supplies the partial space information representing the partial space to the calculation target function selection unit 23, and the process proceeds to step S2.
[0064]
In step S <b> 2, the probability that the calculation target function selection unit 23 is associated with the subspace represented by the subspace information from the subspace detection unit 21 in the calculation target function table stored in the calculation target function table storage unit 24. From the density function or the like, one or more probability density functions or the like are selected based on the speed / accuracy setting information from the speed / accuracy setting unit 27 as necessary, and selection information representing the selected probability density function or the like is selected. This is supplied to the score calculation unit 25.
[0065]
The score calculation unit 25 defines the probability density that defines the acoustic model of the acoustic model database 5 represented by the selection information supplied from the calculation target function selection unit 23 for the words stored in the word dictionary of the dictionary database 6 in step S3. Using function etc., feature vector x_tOutput probability b_s(X_t) And the acoustic score is obtained based on the output probability, and the language score is obtained by referring to the grammar rules of the grammar database 7. Further, the score calculation unit 25 generates a hypothesis (speech recognition result candidate) as necessary based on the acoustic score and the language score, and proceeds to step S4.
[0066]
In step S4, it is determined whether or not the calculation of the acoustic score and the language score has been completed up to the end point of the speech section in which the user has spoken. If it is determined that the calculation has not ended, the process returns to step S1 and Hereinafter, the same processing is repeated for the feature vector. The calculation of the acoustic score and the language score is performed while branching by the beam search method as necessary.
[0067]
If it is determined in step S4 that the calculation of the acoustic score and the language score has been completed up to the end point of the speech section in which the user uttered, the process proceeds to step S5, and the output selection unit 26 determines one or more hypotheses The final score is obtained by comprehensively evaluating the acoustic score and the language score obtained with respect to, for example, a hypothesis that enlarges the final score is selected, output as a speech recognition result, and the matching process is terminated.
[0068]
Next, the calculation target function table stored in the calculation target function table storage unit 24 of FIG. 4 will be described with reference to FIGS. In the following, for example, the feature vector space is 512 subspaces Y₀To Y₅₁₁The HMM as an acoustic model stored in the acoustic model database 5 is 16 probability density functions g₀() To g₁₅Shall be defined in parentheses.
[0069]
In the matching unit 11 of FIG. 4, basically, as described above, in the calculation target function table stored in the calculation target function table storage unit 24, the partial space represented by the partial space information output by the subspace detection unit 21. (Feature vector x_tOne or more probability density functions or the like are selected based on the speed / accuracy information from the probability density function or the like associated with the subspace to which the feature vector x_tOutput probability b_s(X_t) Is obtained, but this output probability b_s(X_t) Is a feature vector x in the calculation target function table._tIt is also possible to calculate using all the probability density functions associated with the subspace to which.
[0070]
That is, in the matching unit 11, the output probability b_s(X_t), The feature vector x_tCan be calculated using a probability density function or the like associated with the subspace to which the_tIt is also possible to calculate using all the probability density functions associated with the subspace to which.
[0071]
Now, output probability b_s(X_t), The feature vector x_tA mode that is calculated using a probability density function or the like associated with the subspace to which the data belongs is called a selectable mode and an output probability b_s(X_t), The feature vector x_tIf the calculation mode using all probability density functions associated with the subspace to which the data belongs belongs to a non-selectable mode, in the non-selectable mode, for example, a calculation target function table as shown in FIG. Is used.
[0072]
That is, in the calculation target function table of FIG.₀To Y₅₁₁Each has its subspace Y_jFeature vector x belonging to (j = 0, 1,..., 511)_tOutput probability b_s(X_t) List of probability density functions used in the calculation of {g_m} Or a floor value is associated.
[0073]
When the calculation target function table of FIG. 6 is used, the calculation target function selection unit 23 sets the probability density function {g_m} Is associated with the subspace Y_jIs received from the subspace detector 21, the subspace Y is received._jProbability density function {g_m} Select all and its probability density function {g_m} Selection information representing all is supplied to the score calculation unit 25.
[0074]
The score calculation unit 25 uses the probability density function represented by the selection information to output the output probability b_s(X_t) Is calculated. Therefore, in this case, the score calculation unit 25 outputs the output probability b in the same manner as described with reference to FIG._s(X_t) Is calculated.
[0075]
By the way, in the calculation target function table of FIG._jFor feature vector x_tOutput probability b_s(X_t) Probability density function {g_m} May be associated with the floor value.
[0076]
The floor value is the subspace Y with which it is associated_jIs a fixed value representing the minimum value of the output probability of the feature vector belonging to, and the subspace Y associated with the floor value_jThe output probability of the feature vector belonging to is the floor value.
[0077]
In other words, the calculation target function selection unit 23 uses the partial space Y associated with the floor value._jIs received from the subspace detector 21, the subspace Y is received._jIs selected, and selection information representing the floor value is supplied to the score calculation unit 25.
[0078]
In the score calculation unit 25, when the selection information represents a floor value, the floor value is converted into an output probability b._s(X_t).
[0079]
Therefore, in this case, the output probability can be obtained without performing a calculation using the probability density function, so that the amount of calculation can be reduced.
[0080]
That is, in the calculation target function table of FIG._jFor the probability density function {g_m} Is associated with the floor value, and the floor value is associated with the partial space Y with the associated floor value._jFeature vector x belonging to_tOutput probability b_s(X_t) Is not required to calculate the probability density function, and as shown in FIG. 2, when using the calculation target function table in which the probability density function is associated with all the partial spaces. Compared to this, the amount of calculation can be further reduced.
[0081]
In the calculation target function table of FIG._ThreeThe floor value “−30.0” is associated with the floor value, and this floor value is a logarithm of the output probability. The same applies to certain floor values shown in the embodiments of FIGS.
[0082]
In the non-selectable mode, a calculation target function table as shown in FIG. 7 can be used in addition to the calculation target function table of FIG.
[0083]
That is, the calculation target function table of FIG. 7 is different from the calculation target function table of FIG._jThe number of probability density functions associated with is added. When the calculation target function table of FIG. 7 is used, the subspace Y to which the feature vector belongs is included in the selection information output from the calculation target function selection unit 23._jIn this case, the score calculation unit 25 can immediately recognize the number of probability density functions that must be calculated to obtain the output probability.
[0084]
Note that the probability density function registered in the calculation target function table has a large contribution to the output probability. Therefore, in the calculation target function table of FIG. 6 or FIG._ThreeOn the other hand, the probability density function is not registered because the probability density function g that defines the acoustic model is not registered.₀() To g₁₅() Each subspace Y_ThreeThis is because the degree of contribution to the output probability of feature vectors belonging to is relatively different. Also, the partial space Y_ThreeFor any feature vector belonging to, the 16 probability density functions g defining the acoustic model₀() To g₁₅The output probability calculated by using () (logarithm thereof) is about −30. Therefore, even when the output probability is set to a fixed value of −30.0, accurate approximation is possible. Subspace Y of the calculation target function table of FIG. 6 and FIG._ThreeFor, the output probability is a fixed value of -30.0.
[0085]
Next, FIG. 8 shows a configuration example of an embodiment of a calculation target function table stored in the calculation target function table storage unit 24 in the selectable mode.
[0086]
In the calculation target function table of FIG._jAnd probability density function {g_m} Or the floor value, a plurality of probability density functions (hereinafter referred to as “calculated number” as appropriate) used for calculating the output probability (and hence the acoustic score) are associated with each other.
[0087]
That is, in the embodiment of FIG. 8, for example, the partial space Y₀, The floor value “−29.0”, the probability density function {g_Five(), G₁(), G₁₅(), G₀()} And the calculated number {0, 1, 4} are associated with each other. For example, the partial space Y₁, Floor value “−45.0”, probability density function {g₀(), G₁(), G₁₇(), G₈(), G_Three(), G_Ten()} And the calculated number {0, 3, 6} are associated with each other. Furthermore, the subspace Y₂For floor value “−20.0”, probability density function {g₂(), G₆(), G_Four()} And the calculated number {0, 3, 3} are associated with each other. Also, the partial space Y_ThreeAre associated with the floor value “−30.0” and the calculated number {0, 0, 0}. In the same manner, the subspace Y_FourTo Y₅₁₀Also, the floor value or probability density function {g_m} Is associated with the calculated number, and the last subspace Y₅₁₁For floor value “−40.0”, probability density function {g₁₅()} And the calculated number {0, 0, 1} are associated with each other.
[0088]
When the calculation target function table of FIG. 8 is used, the calculation target function selection unit 23 first calculates the vector x in the calculation target function table based on the subspace information supplied from the subspace detection unit 21._tSubspace Y to which_jSelect the entry (row). Now, in this way, the subspace Y selected from the calculation target function table based on the subspace information._j, The calculation target function selecting unit 23 selects the one corresponding to the speed / accuracy information supplied from the speed / accuracy setting unit 27 from the plurality of calculated numbers in the selected entry. To do.
[0089]
That is, each of the plurality of calculation numbers in each entry of the calculation target function table represents the number of probability density functions to be selected by the calculation target function selection unit 23, and is registered based on the speed or accuracy required for the speech recognition processing. Has been.
[0090]
Specifically, for example, if it is now possible to set three speeds or accuracy of “high speed / low accuracy”, “medium speed / medium accuracy”, and “low speed / high accuracy” for voice recognition processing, In the embodiment of FIG. 8, three calculation numbers are registered in each entry of the calculation target function table. Of these three calculation numbers, the leftmost calculation number is “high speed / low accuracy”. When the speed or accuracy is set, the second calculation number from the left is “medium speed / medium accuracy”. When the speed or accuracy is set, the rightmost calculation number is “low speed / high accuracy”. Is selected when the speed or accuracy is set.
[0091]
Therefore, in the embodiment of FIG. 8, when the speed / accuracy information supplied from the speed / accuracy setting unit 27 represents “high speed / low accuracy”, the calculation target function selection unit 23 uses the subspace Y.₀To Y₅₁₁Of the three calculated numbers registered in each of the entries, 0, 0, 0, 0,. When the speed / accuracy information represents “medium speed / medium precision”, the calculation target function selector 23 selects the subspace Y₀To Y₅₁₁Of the three calculated numbers registered in each of the entries, 1, 3, 3, 0,..., 0 that are second from the left are selected. Furthermore, when the speed / accuracy information represents “low speed / high precision”, the calculation target function selection unit 23 uses the subspace Y₀To Y₅₁₁Of the three calculated numbers registered in each of the entries, 4, 6, 3, 0,.
[0092]
From the above, the feature vector x_tFor example, subspace Y₀The calculation target function selection unit 23 determines that the subspace Y₀Are selected entries. Further, when the speed / accuracy information supplied from the speed / accuracy setting unit 27 indicates “high speed / low accuracy”, the calculation target function selection unit 23 sets the three calculation numbers “ The leftmost “0” of “0, 1, 4” is selected. In addition, when the speed / accuracy information indicates “medium speed / medium precision”, the calculation target function selection unit 23 selects the left of the three calculation numbers “0, 1, 4” registered in the selection entry. When the second “1” is selected and the speed / accuracy information indicates “low speed / high precision”, the most of the three calculated numbers “0, 1, 4” registered in the selected entry Select “4” on the right.
[0093]
Now, assuming that a selection selected from a plurality of calculation numbers registered in the selection entry as described above is referred to as a selection calculation number, the calculation target function selection unit 23 selects only the number of selection calculations from the selection entry. Select the probability density function.
[0094]
Thus, for example, in FIG.₀When the selected calculation number is “0”, “1”, “4”, the calculation target function selection unit 23 sets the sub-space Y₀0, 1, 4 are selected from the probability density functions registered in.
[0095]
Here, the selection of the probability density function from the selection entry is performed as follows.
[0096]
That is, when the selection calculation number is “0”, the floor value is selected from the selection entry without selecting the probability density function. If the selected calculation number is a value other than “0”, the probability corresponding to the selected calculation number from the left of one or more probability density functions registered there is selected from the selection entry. A density function is selected.
[0097]
Therefore, in FIG.₀When the selected calculation number is “0”, the calculation target function selection unit 23 selects the subspace Y.₀The floor value “−29.0” registered in the entry is selected. When the selected calculation number is “1”, the calculation target function selection unit 23 sets the subspace Y₀Probability density function {g_Five(), G₁(), G₁₅(), G₀()}, Only one from the left, ie {g_FiveSelect ()}. Furthermore, when the number of selected calculations is “4”, the calculation target function selection unit 23 sets the subspace Y₀Probability density function {g_Five(), G₁(), G₁₅(), G₀()} From the left, that is, subspace Y₀All probability density functions registered in {g_Five(), G₁(), G₁₅(), G₀Select ()}.
[0098]
Then, the calculation target function selection unit 23 supplies selection information representing the selected floor value or probability density function to the score calculation unit 25.
[0099]
Here, in the calculation target function table of FIG._ThreeThe three calculated numbers in the entry are all “0”. Therefore, the subspace Y_ThreeWhen the entry is selected, the speed / accuracy information indicates any of “high speed / low precision”, “medium speed / medium precision”, and “low speed / high precision”. The number of selected calculations is “0”. Therefore, the calculation target function selection unit 23 selects the floor value “−30.0”.
[0100]
From the above, the feature vector x_tFor example, subspace Y₀In the “high speed / low accuracy” mode in which the speed / accuracy information is set to “high speed / low accuracy”, the score calculation unit 25 uses the feature vector x_tOutput probability b_s(X_t) Is the floor value “−29.0”. Therefore, in this case, the output probability b_s(X_t) Is obtained without calculation using the probability density function, the accuracy is reduced, but high-speed processing is possible.
[0101]
In the “medium speed / medium precision” mode in which the speed / accuracy information is set to “medium speed / medium precision”, the score calculation unit 25 uses the feature vector x_tOutput probability b_s(X_t) Is a probability density function g of 1_Five(X_t) And the weighting value c based on the above equation (1)._Fiveg_Five(X_t) Is calculated. Therefore, in this case, the output probability b_s(X_t) Is one probability density function g_Five(X_t), The processing speed is reduced but the accuracy is improved as compared with the case of the “high speed / low accuracy” mode.
[0102]
Further, in the “low speed / high precision” mode in which the speed / precision information is set to “low speed / high precision”, the score calculation unit 25 uses the feature vector x_tOutput probability b_s(X_t) Is the four probability density functions g_Five(X_t), G₁(X_t), G₁₅(X_t), G₀(X_t) And the weighted sum c based on equation (1) above._Fiveg_Five(X_t) + C₁g₁(X_t) + C₁₅g₁₅(X_t) + C₀g₀(X_t) Is calculated. Therefore, in this case, the output probability b_s(X_t) Is the four probability density functions g_Five(X_t), G₁(X_t), G₁₅(X_t), G₀(X_t), The processing speed is further reduced as compared with the “high speed / low accuracy” mode, but the accuracy is further improved.
[0103]
In the matching unit 11 shown in FIG. 4, the speed / accuracy information is set according to the operation lever 28 operated by the user. Therefore, the voice recognition process with the speed and accuracy according to the user's request is performed. It becomes possible.
[0104]
When a plurality of probability density functions are registered in the entry of the calculation target function table, the plurality of probability density functions g_m() May be arranged in ascending or descending order of the suffix m, for example, but in the embodiment of FIG. 8, a plurality of probability density functions registered in the entry of the calculation target function table are output probabilities (and eventually , Acoustic score) in descending order of contribution (the leftmost probability density function has the largest contribution to the output probability).
[0105]
Accordingly, in this case, the calculation target function selection unit 23 preferentially selects a probability density function having a large contribution to the output probability (and consequently the acoustic score), and the score calculation unit 25 also has such a probability. Since the density function is preferentially used and the output probability (and hence the acoustic score) is calculated, the output probability generated by omitting calculation of some probability density functions based on the calculation target function table ( The error of the acoustic score can be minimized.
[0106]
In the embodiment of FIG. 8, each entry of the calculation target function table includes three speeds / accuracy of “high speed / low accuracy” mode, “medium speed / medium accuracy” mode, and “low speed / high accuracy” mode. Three calculation numbers for each mode are registered, but each entry in the calculation target function table sets the total number of probability density functions registered in the entry, not the calculation number, and selects the calculation target function. In the unit 23, based on the speed / accuracy information, it is possible to select the number of calculations from an integer value in a range from 0 to the total number of probability density functions (hereinafter referred to as a selection range as appropriate).
[0107]
That is, for example, the subspace Y of the calculation target function table of FIG.₀Entry has four probability density functions {g_Five(), G₁(), G₁₅(), G₀Since ()} is registered, the selection range is 0 to 4, and the integer value in the selection range can be five, 0, 1, 2, 3, and 4. Therefore, the number of calculations is 5 Selected from two integer values.
[0108]
In this case, the movable range of the operation lever 28 in FIG. 4 is divided into five ranges, ie, near the left end, near the center between the left end and the center, near the center, near the center between the center and the right end, and near the right end. When positioned in each of the five ranges, the calculation target function selection unit 23 can select 0, 1, 2, 3, 4 as the number of calculations.
[0109]
In this case, in the score calculation unit 25, when the operation lever 28 is located near the left end, four probability density functions {g_Five(), G₁(), G₁₅(), G₀The output probability is obtained using 0 of ()}, that is, the floor value. When the operation lever 28 is located near the middle between the left end and the center, four probability density functions {g_Five(), G₁(), G₁₅(), G₀()} One probability density function g having the highest contribution to the output probability_FiveBy calculating (), the output probability is obtained. Further, when the control lever 28 is located near the center, four probability density functions {g_Five(), G₁(), G₁₅(), G₀()} Probability density function g having the highest contribution to the output probability_Five() And second highest probability density function g₁By calculating the two of (), the output probability is obtained. Further, when the operation lever 28 is positioned near the middle between the center and the right end, four probability density functions {g_Five(), G₁(), G₁₅(), G₀()} Of three probability density functions g in descending order of contribution to the output probability_Five(), G₁(), G₁₅By calculating (), the output probability is obtained. Further, when the operation lever 28 is positioned near the right end, four probability density functions g_Five(), G₁(), G₁₅(), G₀() By calculating all, the output probability is obtained.
[0110]
Therefore, in this case, speech recognition processing can be performed at five levels of speed or accuracy.
[0111]
As described above, when a probability density function used for calculating the output probability is selected in accordance with the position of the operation lever 28, an acoustic model is defined in each subspace entry of the calculation target function table. 16 probability density function g₀() To g₁₅() All can be registered in descending order of contribution.
[0112]
Next, FIG. 9 shows a configuration example of another embodiment of the calculation target function table stored in the calculation target function table storage unit 24 in the selectable mode.
[0113]
In the embodiment of FIG. 9, the calculation target function table storage unit 24 includes a high speed / low accuracy calculation target function table (FIG. 9A) and a low speed / high accuracy calculation target function table (FIG. 9 ( B)) two calculation target function tables are stored in the calculation target function table storage unit 24, and the calculation target function selection unit 23 receives the speed / accuracy supplied from the speed / accuracy setting unit 27. Based on the information, select either the high-speed / low-precision calculation target function table or the low-speed / high-precision calculation target function table, refer to the selected calculation target function table, and select the probability density function or floor Select a value.
[0114]
That is, when the operation lever 28 is located on the left side, the speed / accuracy setting unit 27 sets that low-speed or high-accuracy voice recognition processing is performed, and the speed / accuracy information to that effect is used as the calculation target function selection unit. 23. In this case, the calculation target function selection unit 23 selects the calculation target function table for low speed / high accuracy (FIG. 9B). On the other hand, when the operation lever 28 is positioned on the right side, the speed / accuracy setting unit 27 sets that high-speed or low-accuracy voice recognition processing is performed, and the speed / accuracy information to that effect is used as the calculation target function selection unit. 23. In this case, the calculation target function selection unit 23 selects the calculation target function table for high speed / low accuracy (FIG. 9A).
[0115]
In the embodiment of FIG. 9, the probability density function registered in each entry in the calculation target function table for low speed / high accuracy (FIG. 9B) is basically the calculation target function table for high speed / low accuracy. In FIG. 9A, a probability density function of 0 or more is added to the probability density function registered in the corresponding entry.
[0116]
Therefore, when a probability density function or the like (probability density function or floor value) is selected with reference to the calculation target function table for low speed / high accuracy, the probability is calculated with reference to the calculation target function table for high speed / low accuracy. Compared with the case where a density function or the like is selected, the amount of calculation required for calculating the output probability is increased, so that the processing speed is reduced but a highly accurate speech recognition result is obtained.
[0117]
Conversely, when selecting a probability density function by referring to the calculation target function table for high speed / low accuracy, select the probability density function, etc., referring to the calculation target function table for low speed / high accuracy. Compared to the case, the accuracy may be deteriorated, but the amount of calculation required for calculating the output probability is reduced, so that the processing speed can be increased.
[0118]
In the embodiment of FIG. 9 as well, the probability density function can be registered in the calculation target function table in descending order of contribution to the output probability, and further, the calculation target is determined according to the position of the operation lever 28. The number of probability density functions selected from the function table can be changed.
[0119]
That is, for example, when the operation lever 28 is located on the left side, the calculation target function table for low speed / high accuracy (FIG. 9B) is selected, and further, depending on how far the operation lever 28 is located on the left side. The number of probability density functions selected from the calculation target function table for low speed / high accuracy can be changed. Further, when the operation lever 28 is located on the right side, the calculation target function table for high speed / low accuracy (FIG. 9A) is selected, and the high speed depends on how far the operation lever 28 is located on the right side. / The number of probability density functions selected from the low accuracy calculation target function table can be changed. In this case, finer control can be performed on the speed or accuracy of the speech recognition process.
[0120]
In the embodiment of FIG. 9, the calculation target function table storage unit 24 stores two calculation target function tables. However, the calculation target function table storage unit 24 performs other speech recognition processing. Three or more calculation target function tables having different numbers of registered probability density functions are stored according to the required speed or accuracy, and the calculation target function selection unit 23 stores the three or more calculation target functions. It is possible to select a calculation target function table to be referenced from the table based on the speed / accuracy information.
[0121]
Next, the series of processes described above can be performed by hardware or software. When a series of processing is performed by software, a program constituting the software is installed in a general-purpose computer or the like.
[0122]
Therefore, FIG. 10 shows a configuration example of an embodiment of a computer in which a program for executing the series of processes described above is installed.
[0123]
The program can be recorded in advance in a hard disk 105 or a ROM 103 as a recording medium built in the computer.
[0124]
Alternatively, the program is stored temporarily on a removable recording medium 111 such as a flexible disk, a CD-ROM (Compact Disc Read Only Memory), an MO (Magneto optical) disk, a DVD (Digital Versatile Disc), a magnetic disk, or a semiconductor memory. It can be stored permanently (recorded). Such a removable recording medium 111 can be provided as so-called package software.
[0125]
The program is installed in the computer from the removable recording medium 111 as described above, or transferred from the download site to the computer wirelessly via a digital satellite broadcasting artificial satellite, LAN (Local Area Network), The program can be transferred to a computer via a network such as the Internet, and the computer can receive the program transferred in this way by the communication unit 108 and install it in the built-in hard disk 105.
[0126]
The computer includes a CPU (Central Processing Unit) 102. An input / output interface 110 is connected to the CPU 102 via the bus 101, and the CPU 102 operates an input unit 107 including a keyboard, a mouse, a microphone, and the like by the user via the input / output interface 110. When a command is input as a result, the program stored in a ROM (Read Only Memory) 103 is executed accordingly. Alternatively, the CPU 102 also transfers from a program stored in the hard disk 105, a program transferred from a satellite or a network, received by the communication unit 108 and installed in the hard disk 105, or a removable recording medium 111 attached to the drive 109. The program read and installed in the hard disk 105 is loaded into a RAM (Random Access Memory) 104 and executed. Thus, the CPU 102 performs processing according to the above-described flowchart or processing performed by the configuration of the above-described block diagram. Then, the CPU 102 outputs the processing result from the output unit 106 configured with an LCD (Liquid Crystal Display), a speaker, or the like via the input / output interface 110, or from the communication unit 108 as necessary. Transmission and further recording on the hard disk 105 are performed.
[0127]
Here, in this specification, the processing steps for describing a program for causing a computer to perform various types of processing do not necessarily have to be processed in time series according to the order described in the flowchart, but in parallel or individually. This includes processing to be executed (for example, parallel processing or processing by an object).
[0128]
Further, the program may be processed by a single computer, or may be processed in a distributed manner by a plurality of computers. Furthermore, the program may be transferred to a remote computer and executed.
[0129]
Note that the speech recognition apparatus shown in FIG. 3 can be applied to, for example, a speech dialogue system or the like when searching a database by voice, operating various devices, or inputting data to each device. It is. More specifically, for example, a database search device that displays map information corresponding to a place name inquiry by voice, an industrial robot that sorts luggage for voice instructions, a voice instead of a keyboard The present invention can be applied to a dictation system that creates text by input, a dialog system in a robot that performs conversation with a user, and the like.
[0130]
In the present embodiment, the speed / accuracy setting unit 27 sets the speed or accuracy of the voice recognition process in accordance with the operation of the operation lever 28 by the user. In addition, for example, it is possible to set based on factors such as resources that can be allocated to the speech recognition processing.
[0131]
That is, for example, when the speech recognition apparatus shown in FIG. 3 is realized by causing a computer as shown in FIG. 10 to execute the program, the CPU 102 generally executes tasks other than the speech recognition processing. Therefore, resources that can be allocated to the speech recognition process change from moment to moment. Therefore, in the speed / accuracy setting unit 27, the CPU 102 recognizes a resource that can be allocated to the speech recognition process, and the speed and accuracy of the speech recognition process are adjusted so that the maximum accuracy can be obtained in real time by the resource. Can be set.
[0132]
In the present embodiment, the score calculation unit 25 calculates the acoustic score based on the continuous HMM method using the continuous amount of feature vectors. However, in the present invention, for example, a discrete-value feature vector is calculated. It is also applicable to the case where the acoustic score based on the discrete HMM method is used.
[0133]
That is, for example, Satoshi Takahashi, Kiyoaki Aikawa, and Shigeki Sagayama. Discrete mixture hmm. In International Conference on Acoustic, Speech, and Signal Processing, pages 971-974, 1997 etc. Is described. According to this discrete mixed distribution type HMM, the feature vector x_tOutput probability b_s(X_t) Is calculated according to the following equation, for example.
[0134]
b_s(X_t) = ΣC_m× G_m(Z_i) ... (2)
[0135]
Here, in Equation (2), C_mIs the mth function G_mIs a weighting factor for () and the function G_m() Is the mth probability function constituting the discrete mixed distribution type HMM. Σ represents a summation for the variable m. Z_iIs the feature vector x_tRepresents a subspace of the feature vector space to which the_tIs obtained by vector quantization. Therefore, Z_iIs the feature vector x_tCan be considered to represent a code vector (representative vector) of a subspace of the feature vector space to which the value belongs, and its value is a discrete value.
[0136]
Since the equation (2) has the same form as the equation (1) described above, even when the discrete mixed distribution type HMM is used, the calculation of the probability function for obtaining the output probability by the calculation object function table is performed. Therefore, the probability density function g in the calculation target function tables shown in FIGS._m() Is a probability function G_mBy using the calculation target function table replaced with (), it becomes possible to perform speech recognition processing with speed and accuracy in accordance with the user's request.
[0137]
In this embodiment, voice recognition based on the HMM method is performed. However, the present invention can also be applied to voice recognition based on other algorithms.
[0138]
Furthermore, in the present embodiment, the calculation target function table is configured using a necessary floor value in addition to the probability density function, but the calculation target function table may be configured without using the floor value. Is possible.
[0139]
The calculation target function tables shown in FIGS. 6 to 9 can be created in principle as follows, for example. That is, each subspace Y of the feature vector space₀To Y₅₁₁Code vector v₀Thru v₅₁₁16 probability density functions g each defining an acoustic model₀() To g₁₅() Obtain the probability output from each, select the probability density function with a relatively large contribution for each subspace as the contribution to the output probability, and register it in the entry of that subspace Thus, a calculation target function table can be created. As the floor value of the calculation target function table, for example, an arbitrary feature vector x belonging to each partial space_tFor the above, it is possible to adopt the minimum value, maximum value, average value, etc. of the output probability calculated according to the equation (1).
[0140]
【The invention's effect】
  Of the present inventionAccording to one aspect,For example, it is possible to perform voice recognition processing with speed and accuracy according to a request from a user or the like. That is, it is possible to perform voice recognition processing that emphasizes speed and voice recognition processing that emphasizes accuracy.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating an exemplary configuration of a conventional speech recognition apparatus.
FIG. 2 is a diagram showing a calculation target function table used in a conventional speech recognition apparatus.
FIG. 3 is a block diagram showing a configuration example of an embodiment of a speech recognition apparatus to which the present invention is applied.
4 is a block diagram illustrating a configuration example of a matching unit 11. FIG.
FIG. 5 is a flowchart illustrating a matching process performed by a matching unit 11;
6 is a diagram showing a configuration example of a first embodiment of a calculation target function table stored in a calculation target function table storage unit 24. FIG.
FIG. 7 is a diagram illustrating a configuration example of a second embodiment of a calculation target function table stored in a calculation target function table storage unit 24;
FIG. 8 is a diagram illustrating a configuration example of a third embodiment of a calculation target function table stored in a calculation target function table storage unit 24;
FIG. 9 is a diagram illustrating a configuration example of a fourth embodiment of a calculation target function table stored in a calculation target function table storage unit 24;
FIG. 10 is a block diagram illustrating a configuration example of an embodiment of a computer to which the present invention has been applied.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Microphone, 2 AD conversion part, 3 Feature extraction part, 5 Acoustic model database, 6 Dictionary database, 7 Grammar database, 11 Matching part, 21 Subspace detection part, 22 Subspace data storage part, 23 Calculation object function selection part, 24 calculation target function table storage unit, 25 score calculation unit, 26 output selection unit, 27 speed / accuracy setting unit, 28 operation lever, 101 bus, 102 CPU, 103 ROM, 104 RAM, 105 hard disk, 106 output unit, 107 input Part, 108 communication part, 109 drive, 110 input / output interface, 111 removable recording medium

Claims (8)

A speech recognition device for recognizing speech,
Extracting means for extracting the feature amount of the voice;
Detecting means for detecting a partial space in which the feature amount of the voice belongs in the feature amount space;
Storage means for storing in association with one or more definition information defining an HMM (Hidden Markov Model) used for matching processing with the voice feature quantity for each of a plurality of partial spaces of the feature quantity space; ,
Selecting means for selecting any one or more definition information from the one or more definition information associated with the partial space to which the audio feature amount belongs;
Using the definition information selected by the selection unit, a matching process between the speech feature quantity and the HMM is performed to obtain a score representing the likelihood that the speech corresponds to the HMM. And a matching means for outputting a speech recognition result of the speech based on the score,
The definition information is a probability density function or a probability function used for obtaining an output probability that the HMM outputs the feature value,
The selection unit is configured to select the voice based on the speed or accuracy of voice recognition processing set according to a user operation or the speed or accuracy of voice recognition processing set according to resources that can be allocated to the voice recognition processing. A speech recognition apparatus that selects a number of the definition information corresponding to the speed or accuracy of the recognition processing in an order corresponding to the magnitude of the definition information contributing to the score. The speech recognition apparatus according to claim 1, wherein the definition information further includes a fixed value representing an output probability that the HMM outputs the feature amount. The storage means also stores the number of the definition information to be selected by the selection means in accordance with the speed or accuracy of the voice recognition processing,
The speech recognition apparatus according to claim 1, wherein the selection unit selects a number of the definition information corresponding to the speed or accuracy of the speech recognition processing. The storage means includes a definition information table in which one or more definition information defining the HMM used for matching processing with the feature amount of speech is associated with each of a plurality of partial spaces of the feature amount space, It is stored for each speed or accuracy of the voice recognition process.
The speech recognition apparatus according to claim 1, wherein the selection unit selects the definition information from the definition information table corresponding to the speed or accuracy of the speech recognition processing. The speech recognition apparatus according to claim 1, wherein the matching unit performs the matching process based on a continuous HMM method or a discrete HMM method using the feature amount of a continuous amount or a discrete value. A speech recognition method for a speech recognition device that recognizes speech,
An extraction step in which the speech recognition apparatus extracts a feature amount of the speech;
A detection step in which the speech recognition device detects a partial space to which the feature amount of the speech belongs in the feature amount space;
The speech recognition apparatus stores one or more definition information defining an HMM (Hidden Markov Model) used for matching processing with the feature amount of speech for each of a plurality of partial spaces of the feature amount space in association with each other. A selection step of selecting any one or more pieces of definition information from the one or more pieces of definition information associated with the partial space to which the audio feature amount belongs in the storage means
The speech recognizer corresponds to the HMM by performing matching processing between the feature amount of the speech and the HMM using the definition information selected in the selection step. A matching step for obtaining a score representing likelihood and outputting a speech recognition result of the speech based on the score;
The definition information is a probability density function or a probability function used for obtaining an output probability that the HMM outputs the feature value,
In the selection step, based on the speed or accuracy of the voice recognition process set according to the user's operation or the speed or accuracy of the voice recognition process set according to the resources that can be allocated to the voice recognition process, A speech recognition method, wherein the number of definition information corresponding to the speed or accuracy of recognition processing is selected in the order corresponding to the magnitude of the definition information contributing to the score. A program for causing a computer to perform speech recognition processing for recognizing speech,
Extracting means for extracting the feature amount of the voice;
Detecting means for detecting a partial space in which the feature amount of the voice belongs in the feature amount space;
Storage means for storing in association with one or more definition information defining an HMM (Hidden Markov Model) used for matching processing with the voice feature quantity for each of a plurality of partial spaces of the feature quantity space; ,
Selecting means for selecting any one or more definition information from the one or more definition information associated with the partial space to which the audio feature amount belongs;
Using the definition information selected by the selection unit, a matching process between the speech feature quantity and the HMM is performed to obtain a score representing the likelihood that the speech corresponds to the HMM. And a program for causing a computer to function as a matching means for outputting the speech recognition result of the speech based on the score obtained.
The definition information is a probability density function or a probability function used for obtaining an output probability that the HMM outputs the feature value,
The selection unit is configured to select the voice based on the speed or accuracy of voice recognition processing set according to a user operation or the speed or accuracy of voice recognition processing set according to resources that can be allocated to the voice recognition processing. A program that selects a number of the definition information corresponding to the speed or accuracy of recognition processing in an order corresponding to the magnitude of the definition information contributing to the score. A recording medium on which a program for causing a computer to perform speech recognition processing for recognizing speech is recorded,
Extracting means for extracting the feature amount of the voice;
Detecting means for detecting a partial space in which the feature amount of the voice belongs in the feature amount space;
Storage means for storing in association with one or more definition information defining an HMM (Hidden Markov Model) used for matching processing with the voice feature quantity for each of a plurality of partial spaces of the feature quantity space; ,
Selecting means for selecting any one or more definition information from the one or more definition information associated with the partial space to which the audio feature amount belongs;
Using the definition information selected by the selection unit, a matching process between the speech feature quantity and the HMM is performed to obtain a score representing the likelihood that the speech corresponds to the HMM. And a program for causing a computer to function as a matching means for outputting the speech recognition result of the speech based on the score obtained.
The definition information is a probability density function or a probability function used for obtaining an output probability that the HMM outputs the feature value,
The selection unit is configured to select the voice based on the speed or accuracy of voice recognition processing set according to a user operation or the speed or accuracy of voice recognition processing set according to resources that can be allocated to the voice recognition processing. A recording medium on which a program for selecting the number of definition information corresponding to the speed or accuracy of recognition processing is selected in the order corresponding to the magnitude of the definition information contributing to the score.

Images