JP2010145784A - Voice recognizing device, acoustic model learning apparatus, voice recognizing method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately perform voice recognition using hidden Markov model. <P>SOLUTION: A feature amount extracting means 111 assigns a frame to input voice, and extracts the feature amount for each frame. A cumulative likelihood calculating means 113, using a hidden Markov model, calculates cumulative likelihood for each state in each frame. At this time, normal likelihood operation is performed in a frame showing consonant, and likelihood operation is performed for a group obtained by collecting similar models in a frame showing vowel. After a group providing a maximum cumulative likelihood is determined, the likelihood operation is performed only for the group. Thus, consonant with small individual difference between speakers is separated from vowel with large individual difference, recognition considering individual difference is performed only for vowel, and as a result, the accuracy of voice recognition processing is increased. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition device, an acoustic model learning device, a speech recognition method, and a program, and in particular, a speech recognition device, an acoustic model learning device, and a speech that can perform speech recognition using a hidden Markov model with high accuracy. The present invention relates to a recognition method and a program.

Mechanical speech recognition can also be performed by the following method if only one sound, for example, “A” is recognized.
First, a frame (time window) having a predetermined length is set for the audio signal waveform, and a numerical feature amount is extracted from each frame.
And the feature-value extracted in each flame | frame is compared with the acoustic model which is a standard pattern.
As a result of the comparison, the sound of the acoustic model that matches the feature amount is set as the recognition result.

  For example, there are five Japanese vowels, “A”, “I”, “U”, “E”, and “O”, which are extracted from a waveform that is captured by a microphone and converted into an electrical signal. If the amount matches the acoustic model “A”, “A” is taken as the recognition result.

  Here, the acoustic model is a feature amount prepared in advance such that “A” is the feature amount, and “I” is the feature amount. Such preparation is equivalent to learning of an acoustic model.

  However, if such an acoustic model is a personal recognition device, it is sufficient that the individual learns the device. However, for example, voice recognition can be performed by a telephone in a public institution or by an unspecified person. It may be necessary to recognize the voice of an unknown person, such as the dictation device used.

  Therefore, an acoustic model learned from the voices of as many people as possible is prepared. In that case, “A” is this numerical range, “I” is this numerical range, and so on, and the acoustic model is prepared in the numerical range. For voice recognition, the feature value obtained through the microphone is the numerical range. If it enters, it will be decided as "A", if it enters this numerical range, it will be decided as "I".

  However, when recognizing a voice spoken by a person, the continuation of phonemes is recognized, and the above-described feature amount takes different numerical values depending on the connection with the phonemes before and after the same phoneme “A”. For this reason, the above-described method cannot be used in speech recognition.

  In general, in speech recognition, a phoneme sequence is regarded as a transition from one steady state to another steady state, and this transition is a so-called Markov process. As an acoustic model, a “Hidden Malkov Model” (hereinafter “Hidden Malkov Model”) The word from which the signal is output is stochastically estimated from the speech signal by a statistical method using “HMM”.

  In this method, a likelihood indicating which feature quantity corresponding to which HMM is output with the highest probability is calculated, and a word corresponding to the HMM having the maximum probability is output as a speech recognition result. Such a speech recognition method is disclosed in, for example, Patent Document 1.

  This likelihood calculation is calculated | required by calculating the following numerical formula (1) of Gaussian distribution, for example.

P _m (Y; μ _m , Σ _m )
= {1 / √ ((2π ) n | Σ j |)} exp (-1/2 (y t -μ t) T Σ -1 (y t -μ t))
J: Number of states t: time
(1)

  Then, the cumulative likelihood value is updated by a Viterbi algorithm in which the calculated likelihood for each HMM is accumulated with respect to the maximum value of the cumulative likelihood value calculated in the previous frame.

  In speech recognition that performs the above-described calculation, the HMM is created by learning from a large amount of utterance data. In particular, in speech recognition for unspecified speakers, utterance data is collected widely for age groups, genders, and the like. As a result, the speech of any person can be recognized.

  However, while the speech of any person can be recognized, the range that can be taken by the HMM values for each phoneme (actually, the range of a multidimensional vector) is expanded, and as a result, the recognition accuracy may be reduced.

JP 2001-356790 A

  The present invention has been made in view of the above circumstances, and an object thereof is to perform speech recognition using a hidden Markov model (HMM) with high accuracy.

In order to achieve the above object, a speech recognition apparatus according to the first aspect of the present invention provides:
A storage unit that stores an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
Probability of calculating the state transition probability of each phoneme for the input speech based on the feature quantity extracted for each of the plurality of predetermined long frames for the input speech and each acoustic model stored in the storage unit A calculation means;
A likelihood calculating means for accumulating the calculated state transition probabilities and calculating a likelihood for each acoustic model;
A cumulative likelihood calculating means for sequentially calculating a cumulative value of likelihood calculated in a frame before the frame;
Speech recognition means for recognizing the input speech based on the cumulative likelihood calculated by the cumulative likelihood calculation means;
It is provided with.

  The reason why the acoustic models are divided in this way is as follows. First, the consonant recognition and the vowel recognition are divided because the consonant has a small individual difference among speakers, whereas the vowel has a large individual difference due to the influence of the vocal cords. The reason why the acoustic model for vowel recognition is divided into a plurality is to deal with individual differences in vowels.

In the above speech recognition apparatus,
Frame identification means for determining whether the voice of each frame is a vowel or a consonant;
Group determination means for determining a group that has learned the acoustic model for vowel recognition when the input speech is a vowel;
It is desirable to provide.

  This is because after a predetermined number of vowels have been recognized, it is desirable to determine a group and perform efficient recognition processing.

In order to achieve the above object, an acoustic model learning device according to the second aspect of the present invention provides:
A storage unit that stores an acoustic model for consonant recognition that learns from all speech data, and an acoustic model for vowel recognition for each group that learns from speech data for each group,
A group number specifying means for specifying the number of groups of an acoustic model for vowel recognition;
Distance calculating means for calculating a distance between groups of the acoustic model for vowel recognition;
Grouping means for making two groups of the shortest distance into one group;
A group number determination means for determining whether the total number of groups is equal to or less than a specified number;
It is provided with.

In order to achieve the above object, a speech recognition method according to a third aspect of the present invention includes:
A speech recognition method for improving accuracy of speech recognition using an acoustic model by a predetermined device,
A model acquisition step of acquiring an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
A feature amount extraction step for setting a plurality of predetermined length frames at a predetermined cycle for the target speech and extracting a feature amount for each frame;
A probability calculating step of calculating a state transition probability of each phoneme for the target speech based on the feature amount extracted in each frame;
A likelihood calculating step for accumulating the calculated state transition probabilities and calculating a likelihood for each acoustic model;
A cumulative likelihood calculating step for sequentially calculating the cumulative likelihood based on the calculated likelihood for each acoustic model and the maximum likelihood calculated in a frame before the frame;
A speech recognition step for performing speech recognition based on the calculated cumulative likelihood;
It is provided with.

In order to achieve the above object, a program according to the fourth aspect of the present invention provides:
Storing an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
Capture the target voice, set a plurality of predetermined length frames for the captured voice in a predetermined cycle, extract the feature amount for each frame,
Based on the feature amount extracted in each frame, the state transition probability is calculated,
Accumulate the calculated state transition probabilities, calculate the likelihood for each acoustic model,
Based on the calculated likelihood for each acoustic model and the maximum likelihood calculated in a frame before the frame, the cumulative likelihood is sequentially calculated,
Performing speech recognition based on the calculated cumulative likelihood,
It is made to function as a voice recognition device characterized by this.

  According to the present invention, speech recognition using a hidden Markov model (HMM) can be performed with high accuracy.

  Embodiments according to the present invention will be described below with reference to the drawings.

(Embodiment 1)
(Voice recognition device)
FIG. 1 is a block diagram showing a configuration of a speech recognition apparatus according to an embodiment of the present invention. As shown in the figure, the speech recognition apparatus 100 includes a control unit 110, an input control unit 120, an output control unit 130, a program storage unit 140, and a storage unit 150.

  The control unit 110 includes, for example, a CPU (Central Processing Unit) and a predetermined storage device (RAM (Random Access Memory)) serving as a work area, and controls each unit of the speech recognition apparatus 100. Each process to be described later is executed based on a predetermined operation program stored in the program storage unit 140.

  The input control unit 120 includes, for example, an ADC (Analog Digital Converter) that performs sampling such as PCM (Pulse Code Modulation) and the like, and analog audio input from a predetermined input device 12 such as a microphone. Convert the signal to a digital signal.

  The output control unit 130 connects, for example, a predetermined output device 13 such as a speaker or a display device, and outputs a voice recognition result or the like by the control unit 110 from the output device 13.

  The program storage unit 140 includes a predetermined storage device such as a ROM (Read Only Memory), a flash memory, and a hard disk device, for example, and stores various operation programs executed by the control unit 110. The program storage unit 140 stores the following operation programs. Each process of the speech recognition apparatus 100 to be described later is realized by the control unit 110 executing these operation programs.

(1) “feature amount extraction program”: a program for extracting feature amounts (feature parameters) of the audio signal converted by the input control unit 120 (2) “likelihood calculation program”: calculating the likelihood for each frame And a program (3) “voice recognition program” for calculating cumulative likelihood: a program for voice recognition based on the calculated cumulative likelihood and an acoustic model

  As shown in FIG. 2, the control unit 110 executes each program stored in the program storage unit 140, so that the feature amount extraction unit 111, likelihood calculation unit 112, cumulative likelihood calculation unit 113, node It functions as the creation unit 114 and the voice recognition unit 115. FIG. 2 is a functional block diagram schematically showing functions of the control unit 110.

  The feature amount extraction unit 111 sets a plurality of predetermined length frames at a predetermined period for the audio signal converted by the input control unit 120, and extracts a power component (feature amount) for each frame.

  The likelihood calculating unit 112 compares the feature amount extracted for each frame with a hidden Markov model (HMM) stored in an acoustic model storage unit 153 to be described later, thereby performing continuous phoneme recognition for each frame. The state transition probability (likelihood) for each HMM is calculated. Here, a predetermined number of states is predetermined for each phoneme, and the probability of transition from one state of each phoneme to which state is obtained by comparing the acquired feature quantity with the HMM. For example, the phoneme of the word “Hachinohe” is “h, a, ch, i, n, o, h, e”, but when the number of states is “3”, each phoneme is “h1, h2”. , h3 ”,“ a1, a2, a3 ”,“ ch1, ch2, ch3 ”, and so on. In this embodiment, it is assumed that the number of states for each phoneme is “3” and the following processes are performed.

  The cumulative likelihood calculating unit 113 obtains a cumulative value of likelihood for each state in each frame, based on the likelihood calculated by the likelihood calculating unit 112 so far.

  Based on grammatical information stored in a grammar storage unit 154, which will be described later, the node creation unit 114 expands a candidate word acquired from the dictionary storage unit 155, which will be described later, in association with a cumulative likelihood.

  The speech recognition unit 115 acquires and outputs candidate words as speech recognition results based on the cumulative likelihood developed by the node creation unit 114.

  The storage unit 150 includes, for example, a RAM (Random Access Memory), a flash memory, a storage device such as a hard disk device, and the like, and stores various types of information necessary for the speech recognition processing of the speech recognition device 100. As shown in FIG. 3, the storage unit 150 includes a voice storage unit 151, a feature storage unit 152, an acoustic model storage unit 153, a grammar storage unit 154, a dictionary storage unit 155, and a cumulative likelihood storage unit 156.

  The audio storage unit 151 buffers the digital signal converted by the input control unit 120 as needed.

  The feature storage unit 152 stores (develops) information indicating the feature amount for each frame extracted by the feature amount extraction unit 111 (hereinafter referred to as “feature amount data”) as needed.

The acoustic model storage unit 153 stores in advance an acoustic model (phoneme model) obtained by modeling all the phonemes constituting the speech to be recognized for the language supported by the speech recognition apparatus 100. In the present embodiment, a “hidden Markov model” (HMM) is used as the acoustic model. The acoustic model storage unit 153 in the present embodiment stores the HMM for consonant recognition and the HMM for vowel recognition learned from all speech data by dividing the HMM into consonants and vowels. Furthermore, the vowel recognition HMM is composed of a plurality of HMMs learned from voice data for each group.
Hereinafter, this grouping will be described.

  Voices uttered by people contain more individual differences in vowels than consonants. Therefore, HMMs are divided into groups of people for only vowels. Specifically, the distance of the mel frequency cepstrum coefficient (MFCC) is obtained, and persons having a short distance are grouped into groups so that voice recognition can be performed for each group.

Here, the number of dimensions of the MFCC is N. Then, assuming that there are V vowels in the recognition target language, numbers 0 to V-1 are assigned to them. Also, let M be the number of speakers and K _mv be the number of phonemes of the vowel v uttered by the mth person.

Then, the MFCC of the kth vowel v uttered by the mth person is C _{m, k, v} = {cn _{, m, k, v} | n = 0,..., N−1}, and the average value G _{m, v} = {gn _{, m, v} | n = 0,..., N−1} is defined as follows.
g _{n, m, v} = (1 / K _{m, v} ) Σk _{= 0} ^{Km, v−1} c _{n, m, k, v}
(2)

Further, a distance D (m1, m2) between vowels uttered by the m1st person and the m2nd person is defined as follows.
D (m1, m2) = Σ v = 0 V-1 Σ n = 0 N-1 s n (g n, m1, v -g n, m2, v) 2
(3)
Here, s _n is an n-dimensional weighting coefficient of MFCC.

In addition, the group A M _A person collects these distances closer's together, and could and group B of human M _B. In that case, the distance DG (A, B) between the group A and the group B is defined as follows.
DG (A, B) = MAX (D (ma _i , mb _j )) (4)

Here, the range of ma _{i is, {ma i | i = 0} , ..., M A -1} is, _{M A} is a number of people belonging to the group A. The range of mb _j is {mb _j | j = 0,..., M _B −1}, where M _B is the number of people belonging to group B.
By the above calculation, the vowel HMMs are divided into a predetermined number of groups.

There is also a method of using the angle magnitude DA of the n-dimensional vector instead of the distance. This is obtained by the following equation.
DA (m1, m2)
= Σv _{= 0} ^V-1 [{Σn _{= 0} ^N-1 (g _{n, m1, v} × g _{n, m2, v} )} / {√ (Σ _{n = 0} ^N-1 g _{n, m1, v} ² ) √ (Σ _{n = 0} ^N-1 gn _{, m2, v} ² )}
(5)
In a broad sense, this can also be regarded as a distance (broadly defined distance) between the m1st person and the m2nd person.
Here, the description of the acoustic model storage unit 153 is finished, and the description of the next part of the storage unit 150 is started.

  The grammar storage unit 154 stores a file that defines grammar rules for a language supported by the speech recognition apparatus 100.

  The dictionary storage unit 155 stores a word dictionary in which phoneme pattern sequence information is registered for each word in a language supported by the speech recognition apparatus 100.

  The cumulative likelihood storage unit 156 stores cumulative likelihood information indicating the cumulative likelihood calculated by the cumulative likelihood calculating means 113. That is, when the cumulative likelihood calculating unit 113 calculates the cumulative likelihood, the node creating unit 114 develops a cumulative likelihood map as shown in FIG. 4 in the cumulative likelihood storage unit 156. In the example shown in FIG. 4, the cumulative likelihood value is developed for each state number of each frame for the word “kensenuma”. In the present embodiment, the reciprocal of the calculated cumulative likelihood is developed in the cumulative likelihood storage unit 156. Therefore, it shows that likelihood is so large that a numerical value is small among the cumulative likelihoods shown in FIG.

  The operation of the speech recognition apparatus 100 configured as described above will be described below with reference to the drawings. Each operation shown below is realized when the control unit 110 executes any or all of the programs stored in the program storage unit 140 in a timely manner.

  First, an outline of a speech recognition operation (“speech recognition processing”) by the speech recognition apparatus 100 according to the embodiment of the present invention will be described with reference to a flowchart shown in FIG. This “voice recognition processing” is started when a voice is input from the input device 12 of the voice recognition device 100 and the voice signal digitally converted by the input control unit 120 is buffered in the voice storage unit 151. And

  First, the feature quantity extraction unit 111 assigns a frame having a predetermined length to the audio signal buffered in the audio storage unit 151 for each predetermined period, extracts the feature quantity in each frame, and stores the feature quantity data as a feature. The data is stored in the unit 152 (step S501). The “frame number” indicating each frame is assigned from “0”.

  Then, the likelihood calculating means 112 sets an initial value “0” to the frame pointer (f) that designates the frame number (step S502).

  Next, the likelihood calculating unit 112 determines whether or not the likelihood calculation has been performed on the frame immediately before the frame (step S503). Since the 0th frame is designated in step S502, the likelihood calculation in the immediately preceding frame is not performed. Therefore, the process proceeds to step S601 shown in FIG.

  Then, the likelihood calculating unit 112 sets an initial value “0” to the state number pointer (s) indicating the state number in the frame (step S601).

  Further, the likelihood calculating means 112 sets an initial value “0” as the likelihood in the number of states (step S602).

  Next, the likelihood calculating unit 112 performs a probability calculation using the Gaussian distribution stored in the acoustic model storage unit 153 (step S603). This calculation is performed using Equation (1) above, but is actually a mixed Gaussian distribution, and the weighted sum of the normal distribution of Equation (1) is obtained. Then, the likelihood is updated with the probability calculated in step S603 (step S604). Note that the information indicating the calculated probability and likelihood is held in a predetermined storage area such as a work area, for example.

  Then, the likelihood calculating unit 112 determines whether there is a further number of states in the frame (step S605).

  If there is a further number of states in the frame (step S605: No), the state number pointer (s) is incremented by 1 (step S606), and probability calculation and likelihood are performed using a Gaussian distribution corresponding to the next number of states. Update is performed (steps S603 and S604).

  When the probability calculation and the likelihood update in all the number of states are completed (step S605: Yes), the cumulative likelihood calculating unit 113 uses the likelihood calculated in each state of the frame, for example, by Viterbi algorithm. The cumulative likelihood for each state is calculated and updated (step S607), and the node creation unit 114 expands the candidate word and the cumulative likelihood in association with each other.

  When the cumulative likelihood for the frame is updated, the likelihood calculating unit 112 determines whether there is a further frame (step S507). When there is a further frame (step S507: No), the likelihood calculating means 112 increments the frame pointer (f) by 1 (step S508), and the same processing is performed in step S503 and subsequent steps for the next frame.

  As described above, since the likelihood calculation is performed in the first frame (number 0), it is determined in step S503 that “there is a probability calculation in the immediately preceding frame” (step S503: Yes). In this case, the likelihood calculating unit 112 refers to the cumulative likelihood developed in the cumulative likelihood storage unit 156, and specifies the state number having the maximum cumulative likelihood value (step S504). This is in order to determine whether the speech of the part is a consonant or a vowel by examining the model and state number of the part having the maximum cumulative likelihood in each frame.

  In the example of FIG. 4, for example, the maximum cumulative likelihood value in the 19th frame is “4939” (as described above, since the reciprocal of the cumulative likelihood value is taken, the one with the smallest absolute value has the maximum likelihood. The corresponding state number is “k3”. Since “k3” is the third state part of “k” in “Kenuma (k, e, s, e, N, n, u, m, a)”, it is understood that it is a “consonant”.

  In this way, the likelihood calculating means 112 determines whether the sound of the frame is a “consonant” or a vowel (step S505).

If the voice is not a “vowel” (step S505: No), the process proceeds to step S601 shown in FIG.
On the other hand, when the voice is a “vowel” (step S505: Yes), a vowel comparison process is performed in step S506. Hereinafter, this process will be described with reference to FIGS.

  In FIG. 7, the likelihood calculating unit 112 first determines whether or not a group to be compared with the feature amount data has been determined (step S701). This process can be performed, for example, by referring to a flag indicating that the group has been determined.

  When the group has been determined (step S701: Yes), only the vowel HMMs of the determined group among all the vowel HMMs are focused (step S702). This process can be performed, for example, by setting the determined group number in the pointer g (step S702).

  Then, an initial value “0” is set to the state number pointer sg of the corresponding group (step S711), and thereafter, the same processing as that of steps S602 to S607 described above in steps S712 to S717 is performed. HMM is used as a comparison target.

  After the above process is completed, the process returns to step S507 described above and the processes of steps S507 to S509 already described are performed.

  On the other hand, when the group has not been determined in step S701 (step S701: No), the processing shown in FIG. 8 is performed.

  In FIG. 8, first, an initial value “0” is set in a counter for each group that counts the number of times each group is referenced (step S801). Then, the first group number “1” is set in the pointer g to process all the groups as comparison targets (step S802).

  Then, an initial value “0” is set to the state number pointer sg of the corresponding group (step S811), and thereafter, in steps S812 to S817, the same processes as in steps S712 to S717 and steps S602 to S607 described above are performed. The vowel HMM of the first group is used as a comparison target.

  When the processing of the first group is completed, the likelihood calculating unit 112 determines whether there is a further group (step S818). When there is a further group (step S818: No), the likelihood calculating unit 112 increments the group pointer (g) by 1 (step S819), and the same processing is performed in step S811 and subsequent steps for the next group.

  On the other hand, when there is no further group (step S818: Yes), the above process is complete | finished and the counter of the group which produced HMM with the highest probability is incremented by 1 (step S820). Then, it is determined whether or not it is time to determine a group (step S821). This determination can be made, for example, by determining whether or not a vowel comparison process has been performed a predetermined number of times or more.

  As a result of this determination, when it is time to determine a group (step S821: Yes), the group number for which the largest value is indicated in the group counter cnt (g) is set to the determined group (step S822). ). With this setting, in the processing of the next frame, the set number of the determined group is moved to the group pointer (g) in step S702 described above, and vowel comparison processing is performed only for the determined group.

  After the above process is completed, the process returns to the above-described step S507 and the processes of steps S507 and S508 already described are performed. Here, when there is a further frame (step S507: No), the likelihood calculating unit 112 increments the frame pointer (f) by 1 (step S508), and performs the same processing in step S503 and subsequent steps for the next frame. .

  On the other hand, when the frame is the final frame (step S507: Yes), a predetermined output process (step S509) is performed to output the voice recognition result. Here, the speech recognition means 115 refers to the candidate words and the cumulative likelihood expanded in the cumulative likelihood storage unit 156, and traces the node back from the final state of the final frame to output the recognition result (word ) And output as voice or text information by the output control unit 130.

  As described above, the speech is recognized by separating it into consonants and vowels, and the vowels are divided into groups, and the HMM is determined from the determined group, so that the recognition process can be speeded up and the recognition accuracy can be improved. Can be increased.

  By the way, as a method for determining a group of vowels, there are the following methods in addition to the above-described method for determining when a vowel is compared more than a specified number of times. This is performed by executing the processes of steps S830 to S832 shown in FIG. 9 instead of steps S820 to S822 described above.

  In this method, when there is no further group at step S818 (step S818: Yes), the average likelihood of each group is calculated (step S830). This is obtained by summing the likelihood for each group and dividing by the total number of HMMs in each group.

  Then, it is determined whether or not the ratio of the average likelihood between the maximum average likelihood group and the second average likelihood group is equal to or higher than a ratio specified in advance (step S831).

  When this ratio is equal to or higher than a ratio specified in advance (step S831: Yes), the group is determined as the maximum average likelihood group (step S832).

  According to the above processing, there is no need to determine the timing of determination by another factor, no need to count the number of times the group has been adopted, and the average value and the maximum value of the likelihood calculated with speech recognition, etc. By calculating, there is an advantage that the group to be adopted can be determined by itself.

(Embodiment 2)
(Acoustic model learning device)
Next, an apparatus for learning the above-described group-specific HMM will be described.

This apparatus is realized by using the apparatus of FIG. 1 described above.
In FIG. 1, the acoustic model storage unit 153 (see FIG. 3) of the storage unit 150 first stores a normal HMM that cannot be divided into consonants and vowels. The control unit 110 functions as an acoustic model learning device by executing an HMM learning program (not shown) stored in the program storage unit 140.

  As shown in FIG. 10, this apparatus functionally includes a group number designating unit 116, a distance calculating unit 117, a grouping unit 118, a group number determining unit 119, and a vowel consonant sorting unit 1110. Prepare.

  The group number designation means 116 designates the number of vowel recognition HMM groups in a predetermined area provided in the acoustic model storage unit 153 (see FIG. 3) of the storage unit 150. This can be performed by causing the control unit 110 to store the specified value in the storage unit 150 by the user specifying the input device 12 using, for example, a keyboard.

  The distance calculation means 117 calculates the distance between groups of HMMs for vowel recognition in accordance with the above-described mathematical expressions (3) and (4), or mathematical expressions (5) and (4).

  The grouping means 118 sets two groups at the shortest distance among all the groups as one group. This can be done by unifying the different group numbers of the two groups determined to be at the shortest distance into any one number and subtracting the numerical value “1” from the total number of groups.

The group number determination means 119 determines whether the total number of groups has decreased as a result of group unification, and the total number of groups has become equal to or less than the specified number.
The vowel consonant sorting means 1110 sorts whether the HMM stored in a predetermined area of the storage unit 150 is a vowel or a consonant and stores them in separate areas.

Next, the operation of the above-described apparatus will be described with reference to FIG.
It is assumed that the number of groups G specified for defining the total number of vowel HMMs for a person is set in the work area of the storage unit 150 or the like via the input device 12 before the operation is started. .

  First, the control unit 110 collects HMM audio data from the medium on which the voice of the speaker is recorded, and stores it in a predetermined area of the storage unit 150 (step S901). In the initial state, the phoneme HMMs of all speakers are regarded as one group each consisting of one person.

  Then, the vowel consonant sorting unit 1110 of the control unit 110 initializes “0” in each of the person pointer m, the voice data pointer n of each person, the consonant HMM pointer s, and the vowel HMM pointer g (step). S902).

  Next, the vowel consonant selection unit 1110 determines whether the nth data of the mth person is a vowel or a consonant (step S903). For this determination, for example, an HMM for performing only the determination of vowels or consonants is prepared, and the probability of being a vowel and the probability of being a consonant are calculated by likelihood calculation as in the case of the speech recognition described above. This can be done by comparing the two.

  If it is a vowel (step S903: Yes), the HMM is stored in the storage area of the vowel HMM (step S904), and the pointer g of the vowel HMM is incremented by 1 (step S905).

  If it is a consonant (step S903: No), the HMM is stored in the consonant HMM storage area (step S906), and the consonant HMM pointer s is incremented by 1 (step S907).

In this way, when the processing of the first voice data is completed, the vowel consonant sorting unit 1110 determines whether or not there is further voice data (step S921). When there is further voice data (step S921: No), the vowel consonant sorting unit 1110 increments the pointer n of the voice data by 1 (step S922), and the same process is performed on and after the next voice data in step S903.
On the other hand, if there is no further audio data (step S921: Yes), the process proceeds to the next person's processing (steps S903 to S907).
When the processing of the first person is completed, the vowel consonant sorting unit 1110 determines whether there are more persons (step S908). When there is a further person (step S908: No), the vowel consonant selection unit 1110 increments the person's pointer m by 1 (step S909), and the same process is performed on and after the next person at step S903 and thereafter.

On the other hand, when there is no further person (step S908: Yes), it progresses to the next process (step S910).
Subsequently, the group number determination means 119 determines whether or not the total group number m of the person is equal to or less than the specified group number G.

  If the total group number m of the person is not less than the specified group number G (step S910: No), the distance calculation unit 117 performs inter-HMM recognition for vowel recognition according to the above-described equation (3) or equation (5). Is calculated (step S911).

  Subsequently, the grouping unit 118 sets the two vowel recognition HMMs determined to be within the shortest distance as one group (step S912). This can be performed, for example, by unifying the smaller group number assigned to each HMM.

  Then, the grouping means 118 subtracts the numerical value “1” from the total number m of people (step S913). Thereafter, the process returns to step S910 described above.

  If the total group number m of the person is not less than or equal to the designated group number G in step S910 (step S910: No), the distance calculation means 117 may calculate the formula (3) and formula (4) or formula ( 5) and the distance between the HMMs for vowel recognition are calculated according to equation (4) (step S911). Here, there is also a group including two or more voice data due to the above-mentioned one grouping.

  Subsequently, the grouping unit 118 sets two vowel recognition HMMs or groups determined to be at the shortest distance as one group (step S912). This can be performed, for example, by unifying the smaller group number assigned to each HMM.

  Then, the grouping means 118 subtracts the numerical value “1” from the total number m of people (step S913). Thereafter, the process returns to step S910 described above.

  When the above process is repeatedly executed and the total number of groups m of the person is equal to or less than the specified number of groups G (step S910: Yes), the group number determination unit 119 ends all the processes.

  As described above, the storage unit 150 stores the HMM for consonant recognition learned from all speech data, the HMM for vowel recognition for each group of people learned from the speech data for each group of people, Is memorized.

  Note that the speech recognition device 100 and the acoustic model learning device (100) in the above embodiment can be configured with a dedicated device, for example, a general-purpose computer device such as a personal computer. In this case, the speech recognition apparatus 100 according to the present invention can be configured by installing a program for realizing the processing described in the above embodiment on the computer apparatus. The distribution method of the program in this case is arbitrary. For example, the program can be distributed by being stored in a recording medium such as a CD-ROM, and can be distributed via a communication medium such as the Internet by being superimposed on a carrier wave. be able to.

  That is, the speech recognition device according to the present invention can be realized as an application that operates on a personal computer, a game device, or the like in addition to being realized as a portable translation device, for example, and realizes highly accurate speech recognition. Is.

  Further, by adding a program for realizing each process according to the present invention to an existing voice recognition device or voice recognition application (for example, version upgrade), the voice recognition process can be made highly accurate. .

  As described above, according to the present invention, high recognition accuracy in voice recognition can be realized.

It is a block diagram which shows the structure of the speech recognition apparatus which concerns on embodiment of this invention. It is a functional block diagram which shows the function implement | achieved by the control part of the speech recognition apparatus shown in FIG. It is a figure which shows the detail of the memory | storage part shown in FIG. It is a figure which shows the example of the cumulative likelihood value developed by the cumulative likelihood storage part shown in FIG. It is a flowchart for demonstrating the "voice recognition process" which concerns on embodiment of this invention. It is a flowchart for demonstrating the "voice recognition process" which concerns on embodiment of this invention. It is a flowchart for demonstrating the "vowel comparison process" which concerns on embodiment of this invention. It is a flowchart for demonstrating the "vowel comparison process" which concerns on embodiment of this invention. It is a flowchart for demonstrating the "vowel comparison process" which concerns on embodiment of this invention. It is a block diagram which shows the structure of the acoustic model learning apparatus which concerns on embodiment of this invention. It is a flowchart for demonstrating the "HMM learning process" which concerns on embodiment of this invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Voice recognition apparatus (acoustic model learning apparatus), 111 ... Feature-value extraction means, 112 ... Likelihood calculation means, 113 ... Cumulative likelihood calculation means, 114 ... Node creation means, 115 ... Speech recognition means, 116 ... Number of groups Designating means, 117 ... distance calculating means, 118 ... grouping means, 119 ... group number judging means, 1110 ... vowel consonant sorting means, 151 ... speech storage section, 152 ... feature storage section, 153 ... acoustic model storage section, 154 ... Grammar storage unit, 155 ... dictionary storage unit, 156 ... cumulative likelihood storage unit

Claims (5)

A storage unit that stores an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
Probability of calculating the state transition probability of each phoneme for the input speech based on the feature quantity extracted for each of the plurality of predetermined long frames for the input speech and each acoustic model stored in the storage unit A calculation means;
A likelihood calculating means for accumulating the calculated state transition probabilities and calculating a likelihood for each acoustic model;
A cumulative likelihood calculating means for sequentially calculating a cumulative value of likelihood calculated in a frame before the frame;
Speech recognition means for recognizing the input speech based on the cumulative likelihood calculated by the cumulative likelihood calculation means;
A speech recognition apparatus comprising: Frame identification means for determining whether the voice of each frame is a vowel or a consonant;
Group determination means for determining a group that has learned the acoustic model for vowel recognition when the input speech is a vowel;
The speech recognition apparatus according to claim 1, further comprising: A storage unit that stores an acoustic model for consonant recognition that learns from all speech data, and an acoustic model for vowel recognition for each group that learns from speech data for each group,
A group number specifying means for specifying the number of groups of an acoustic model for vowel recognition;
Distance calculating means for calculating a distance between groups of the acoustic model for vowel recognition;
Grouping means for making two groups of the shortest distance into one group;
A group number determination means for determining whether the total number of groups is equal to or less than a specified number;
An acoustic model learning device comprising: A speech recognition method for improving accuracy of speech recognition using an acoustic model by a predetermined device,
A model acquisition step of acquiring an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
A feature amount extraction step for setting a plurality of predetermined length frames at a predetermined cycle for the target speech and extracting a feature amount for each frame;
A probability calculating step of calculating a state transition probability of each phoneme for the target speech based on the feature amount extracted in each frame;
A likelihood calculating step for accumulating the calculated state transition probabilities and calculating a likelihood for each acoustic model;
A cumulative likelihood calculating step for sequentially calculating the cumulative likelihood based on the calculated likelihood for each acoustic model and the maximum likelihood calculated in a frame before the frame;
A speech recognition step for performing speech recognition based on the calculated cumulative likelihood;
A speech recognition method comprising: Storing an acoustic model for consonant recognition learned from all speech data and a plurality of acoustic models for vowel recognition learned from speech data for each group;
Capture the target voice, set a plurality of predetermined length frames for the captured voice in a predetermined cycle, extract the feature amount for each frame,
Based on the feature amount extracted in each frame, the state transition probability is calculated,
Accumulate the calculated state transition probabilities, calculate the likelihood for each acoustic model,
Based on the calculated likelihood for each acoustic model and the maximum likelihood calculated in a frame before the frame, the cumulative likelihood is sequentially calculated,
Performing speech recognition based on the calculated cumulative likelihood,
A program for functioning as a voice recognition device.

Images