JP2000003190A - Voice model learning device and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To eliminate dictionary dependency of discriminative learning and to generate a discriminative learning model with high discriminative performance for sets of all words in a specified language by beforehand storing a pair of monosyllable models different from each other in a dictionary part as a learning word pattern. SOLUTION: A voice analytic part 101 inputs a learning voice to calculate/ output the acoustic feature amount of the learning voice in respective frame at every prescribed time (frame). The dictionary part 102 holds a learning monosyllable pair dictionary (pattern: a pair of monosyllable models and word model constituting it). A pattern matching part 103 inputs the feature amount per frame, and outputs correspondence with the learning voice of all models in the monosyllable dictionary and its likelihood. A calculation part 104 calculates/ outputs correction amounts of parameters of respective models in the dictionary from the correspondence, the likelihood and the monosyllable pair dictionary. A correction part 105 corrects the parameters of respective models in the dictionary by using the correction amounts, and a control part 106 inputs the likelihood to decide a calculation end.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a speech model learning device and a speech recognition device using the same.

[0002]

2. Description of the Related Art A speech model used for speech recognition is often trained by a speech model learning device using actual speech. Such learning methods include a method called a non-discriminating learning method and a method called a discriminating learning method.

As a non-discriminating learning method, EM (Expectation-
Maximization) algorithm. This method is a learning algorithm for forming an HMM (hidden Markov model).

In the HMM, only a symbol string of an input pattern is observed, and a transition between states cannot be directly observed. Therefore, assuming a model, calculating the number of state transitions from observation data (input pattern), and performing the maximum likelihood estimation of the transition probability and the output probability based on the calculated number, to update the parameters of the HMM. An estimation algorithm that repeats these two steps until convergence is the EM algorithm.

In the EM learning method using the EM algorithm, for a certain word, a model is generated such that the speech (usually a plurality of words) that generated the word best fits as a whole. The model's goodness of fit can be measured using a forward-backward score or a viterbi (Vite
rbi) Often represented by scores.

Normally, in the EM learning method, a model is generated for a large number of words such as several hundred words. Then, for subwords (syllables, phonemes, etc.) that are commonly included in a plurality of words, a model is generated that is balanced so as to fit all the words evenly.

The EM algorithm is described in, for example, Kiyohiro Kano, "Basics of Speech Recognition" (May 1, 1996)
7), pp. 52-57.

[0008] On the other hand, the discriminative learning method is based on the word model of the utterance content of a speech (a plurality of utterances) generating a certain word, and the word model (error, Model)
Learning to make a difference. In this method, since the purpose is to make a difference between the word model and the error model, the fitness of each model to the utterance content is not so important, and the fitness of the model indicating the utterance content is the fitness of the error model. Adjust between the two models to be higher (advantageously) higher. Normally, the discriminative learning method is applied to a model obtained by the above-described EM learning or the like, so that the degree of conformity of both models is reduced in most cases. Note that the error model is a word model included in a dictionary, and may be all word models (excluding those representing utterance contents) in the dictionary, or a specific word may be selected.

Such an identification learning method includes an identification error.
Minimization (MCE: Minimum Classification Error) learning
There is a law, for example, in Junichi Takahashi and Shigeki Sagayama
A paper by the author "For small data learning using minimization of identification errors.
Research on the Initial Model in Japan ", Proceedings of the Acoustical Society of Japan,
3-5-8 (March 26, 1996), pp. 123-124
It is described in. This MCE learning method uses class identification
Is the discriminant function g_kAnd for a sample X
Error function d expressed by the difference between_k (X,
Λ), the loss function l (d_k)
Is used to evaluate the effective number of identification errors.
The model parameter Λ with the criterion to minimize
This enhances the model's discrimination ability.
This is the competition between the correct class for a sample X and this.
This is a learning method that takes into account the relationships between the classes to be combined. MCE
The method is expressed by the following formulas 1 to 5 by formulation.
You.

[0010]

(Equation 1)

[0011]

(Equation 2)

[0012]

(Equation 3)

[0013]

(Equation 4)

[0014]

(Equation 5) As shown in Expression 5, in this method, the optimal parameters are obtained while gradually adjusting the model parameters Λ using the steepest descent method. Note that the parameter ε _t controls the learning step size.

[0015]

Generally, a speech model created by discriminative learning has an advantage of higher discrimination performance than a speech model created by non-discriminatory learning.

However, the speech model created by the MCE learning method has a high discriminating ability for words included in the dictionary used at the time of learning, but a combination of other words (words not included in the dictionary) Has a problem that it does not have such a high discrimination performance.

More specifically, the MCE learning method uses a specific word set including a plurality of words as a dictionary used for identification learning. In this case, all the words in the dictionary other than the word corresponding to the learning utterance become the error model. As a result, the ability to identify words in the learning dictionary is improved. However, in such a learning method in which a specific word set is used as a dictionary, even if the misrecognition rate in the dictionary can be improved, the identification performance for a word set outside the dictionary is not guaranteed. As an extreme example, in the case where a specific syllable model does not exist in the dictionary, there is a weakness that the conventional recognition learning does not improve the false recognition rate at all for such a syllable model.

An object of the present invention is to eliminate such a dictionary dependence of discrimination learning, and to provide a speech model learning apparatus capable of creating a discrimination learning model with high discrimination performance for all word sets in a specific language, and to use the speech model learning apparatus. To provide a speech recognition device.

[0019]

According to the present invention, there is provided a speech model learning apparatus comprising: a speech analysis unit for calculating a feature amount of an input speech at fixed time intervals; a dictionary unit for storing a plurality of learning word patterns; A pattern matching unit that performs matching between the feature amount for each fixed time and each of the plurality of learning word patterns, and calculates a correspondence and likelihood with each learning word pattern, based on the correspondence and the likelihood. A calculation unit that calculates a parameter correction amount of each learning word pattern, and a correction unit that corrects each of the plurality of word patterns stored in the dictionary unit based on the parameter correction amount, wherein the dictionary The unit stores a pair of mutually different monosyllable models as the learning word pattern.

Further, the speech model learning apparatus of the present invention determines whether or not to perform the learning process again using the learning word pattern having the parameter corrected by the correction unit based on the likelihood. When performing the process, a control unit is provided for instructing the pattern matching unit to re-execute the learning process.

Further, in the speech model learning apparatus according to the present invention, the dictionary unit holds a pair of word models as the word patterns in place of the pair of mutually different monosyllable models, and one of the pair of word models. Is a word model corresponding to a predetermined input voice, and the other word model is a word model closest to the predetermined input voice obtained by a continuous word recognition algorithm.

Still further, in the speech model learning apparatus according to the present invention, the dictionary unit holds a plurality of pairs of word models as the word patterns instead of the pair of different single syllable models, and one of each pair of the word models. Are word models corresponding to different predetermined input voices, and the other word model of each pair is a word model closest to the predetermined input voice obtained by a continuous word recognition algorithm. And

Further, the speech recognition apparatus of the present invention comprises a speech analysis unit for calculating a feature amount of the input speech at regular intervals, a dictionary unit for storing a pair of word patterns, and the feature at regular intervals. A calculating unit that performs matching between the amount and each of the pair of word patterns and calculates the likelihood of each of the pair of word patterns; and inputting one of the pair of word patterns based on the likelihood into the input voice. And a pair of word patterns stored in the dictionary unit are configured by connecting monosyllable models in utterance order, and each monosyllable model is
It is characterized by being generated by discriminative learning using a pair of learning word patterns paired with a single syllable model of the other word pattern in the same utterance order.

The speech recognition apparatus according to the present invention further comprises a speech analysis unit for calculating a feature amount of the input speech at fixed time intervals, a dictionary unit for storing word patterns corresponding to a plurality of words, A calculating unit that performs matching between the feature amount of each word and each of the word patterns to calculate the likelihood of each word pattern, and obtains the word pattern closer to the input voice from the likelihood of the word pattern. Output unit that outputs a single syllable model that constitutes each word as a word pattern corresponding to the plurality of words. Is generated by discriminative learning using a pair of learning word patterns that form a pair with each of the monosyllable models that compose the other words in the order of utterance. number Characterized in that it was prepared word pattern of the corresponding number.

Further, in the speech model learning method according to the present invention, a first step of calculating a feature amount for each fixed time for an input monosyllable utterance, and the feature amount for each fixed time are stored in advance. Performing a matching with each of a pair of learning single syllable patterns, a second step of determining the correspondence and likelihood between the single syllable utterance and each learning single syllable pattern, and based on the correspondence and likelihood, A third step of calculating a parameter correction amount of each learning single syllable pattern, and a fourth step of correcting each of the plurality of learning single syllable patterns stored in advance based on the parameter correction amount.
And a step of:

In the speech model learning method according to the present invention, after the fourth step, it is determined whether or not to perform the learning process again using the learning word pattern having the corrected parameter based on the likelihood. When the learning process is performed again, the fifth and subsequent steps are repeated.
Characterized by comprising the steps of:

Still further, in the speech model learning method according to the present invention, a first step of calculating a feature amount of the input speech at fixed time intervals, and a pair of words stored in advance with the feature amount at fixed time intervals A second step of performing matching with each of the patterns and calculating the likelihood of each of the pair of word patterns, and determining that one of the pair of word patterns corresponds to the input voice based on the likelihood. And a word model to which a single syllable model obtained by a speech model learning method using a pair of learning single syllable patterns is used as the pair of word patterns. And

[0028]

The speech model learning apparatus identifies only the syllable pair of all syllable pairs appearing in a specific language (for example, Japanese, such as a general language including all Japanese and English). Create a discriminative learning model. That is, this speech model learning device holds only a pair of syllable models as a learning dictionary.

In this speech model learning apparatus, discrimination learning is performed for all syllable pairs. As a result, disadvantages caused by depending on the contents of the dictionary used for recognition learning, such as the absence of a specific syllable model in the dictionary, are eliminated.
However, with this device, a model with high single-syllable identification accuracy is generated, but in order to express a word model represented by continuous syllables, accuracy is slightly higher because vocal deformation between syllables is not expressed. It can be short.

Therefore, a syllable string model corresponding to the input and a model expressing all syllable strings included in a specific language may be used as the learning dictionary.

In this case, by performing discrimination processing on syllable pairs and performing discrimination learning on continuous syllables, a model representing vocal deformation between syllables can be simultaneously learned, and discrimination learning with higher discrimination performance A model can be created.

[0032]

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

FIG. 1 shows a first embodiment of the present invention. This speech model learning device can be used for a specific language (for example, Japanese or English, or a combination thereof).
In any language. This is for discriminative learning of all the syllables that can appear in ()) in pairs with other syllables. For example, in the case of Japanese, if there are 50 syllables, 49 syllable pairs can be considered for each syllable, and for each of the {(50 × 49) / 2} syllable pairs excluding overlapping pairs , To perform identification learning. In other words, in this speech model learning device, the syllables to be treated at one time are a pair of syllables, that is, two syllables, but the discrimination learning is finally performed for all possible pairs.

The speech model learning apparatus shown in FIG. 1 receives a learning speech as an input, calculates an acoustic feature amount of the learning speech in each frame at predetermined time intervals (referred to as frames), and outputs a speech analysis unit 101. Single syllable pair dictionary (pattern:
A dictionary unit 102 that holds a pair of monosyllable models and a subword model that constitutes each monosyllable model), and a feature amount for each frame output from the speech analysis unit 101 as inputs.
The correspondence between the monosyllable pairs from the dictionary unit 102 and the learning speeches of all models in the dictionary and the pattern matching unit 103 that outputs the likelihood, the association and the likelihood output from the pattern matching unit 103, and the dictionary. A calculating unit 104 for calculating and outputting a correction amount of a parameter of each model in the dictionary from the monosyllable pair dictionary of the unit 102; and using the correction amount output from the calculation unit 104 to calculate a parameter of each model in the dictionary. It has a correction unit 105 for correction and a control unit 106 that receives the likelihood output from the pattern matching unit 103 and determines whether the calculation is completed.

Next, the operation of the speech model learning apparatus will be described with reference to FIG. 1 and FIGS.
Note that, as an example, a case where single syllable utterances of “shi” and “chi” are used will be described.

First, in step S1 shown in FIG. 2, the speech analysis unit 101 divides an input speech into frames by a predetermined period, and performs a frequency analysis of the speech in the section for each frame. Then, the feature amount of the input voice is calculated for each frame and output to the pattern matching unit 103. here,
The power of the voice, the power change amount, the cepstrum, the cepstrum change amount, and the like can be used as the feature amount.

The dictionary unit 102 holds two learning single-syllable pair dictionaries (single-syllable models). Here, as shown in FIG. 3, a dictionary expressing a single syllable “shi” and a dictionary expressing a single syllable “chi”. Also, the dictionary unit 102
It has a sub-word dictionary (sub-word model) D1 that includes all sub-words corresponding to “shi” and “chi”. Further, the dictionary unit 102 has a sub-word dictionary D2 for copying the sub-word dictionary D1. This sub-word dictionary D2 has the same contents as the sub-word dictionary D1 in the initial state.

Next, in step S2, the pattern matching unit 103 stores the data stored in the dictionary unit 102 based on the feature amount of each frame input from the speech analysis unit 101.
The subword model corresponding to each of the single syllable models is extracted. Further, the likelihood is obtained for each extracted sub-word model. Pattern matching 1
03 is the obtained correspondence and the extracted subword model,
And its likelihood are output to calculation section 104.

The calculation unit 104 determines in step S3
The correction amount of each parameter in each subword model is obtained, and is output to the correction unit 105 together with the likelihood from the pattern matching unit 103. As a method for obtaining the correction amount, a conventional method, for example, the MCE method can be used (see Expressions 1 to 5).

When the correction unit 105 receives the correction amount from the calculation unit, in step S4, the subword dictionary D1
For each of the models described above, the correction amount is added to the feature amount to make correction. After completing the correction of the feature amount, the correction unit 105 outputs the likelihood from the calculation unit 104 to the control unit 106,
Notifies the end of the correction process.

When the correction process in the correction unit 105 is completed, the control unit 106 calculates the value of the loss function based on the likelihood supplied from the calculation unit 104. Here, the control unit 10
Numeral 6 stores the value of the loss function in the previous processing, and in step S5, compares the calculated value of the loss function with the stored value of the loss function.

If the value of the loss function obtained here is equal to or less than the previous stored loss function value, the control unit 106 newly stores the obtained loss function value and sends it to the dictionary unit 102 for the sub-word dictionary. An instruction is issued to copy D1 to the sub-word dictionary D2. The sub-word dictionary D1 is modified in step S4, and at that time the sub-word dictionary D2 is not modified, so that the contents of the two dictionaries are different, but now the same. The control unit 106
Instructs the pattern matching unit 103 to perform the processing again. In the initial state, the dummy value = 1.0, that is, the upper limit of the loss function value is set as the value of the loss function. Therefore, in the first process, the control unit 106 always performs Works. Thereafter, the pattern matching unit 103, the calculation unit 104, and the correction unit 1
In step 05, the operations from step S2 to step S4 described above are performed again. In addition, the control unit 106 again executes step S5
Then, the calculated value of the loss function is compared with the stored value of the loss function.

On the other hand, if the calculated value of the loss function exceeds the stored value of the previous loss function, the control unit 106
The contents of the sub-word dictionary D2 of the dictionary unit 102 are output as a learning model. That is, the control unit 106 outputs not the subword model corrected by the current process but the subword model corrected by the previous process as a learning model.

For example, as shown in FIG.
The value of the loss function calculated in step is the number of processing times (loop times)
At the first, second, and third times, "0.500",
In the case of “0.450” and “0.480”, 1
At the second time, the obtained value is smaller than the initial value “1.0”, so that the second processing is performed. In the second time, the calculated value “0.
Since “450” is smaller than the previous value “0.500”, the third processing is performed, but the value “0.480” obtained the third time is larger than the previous value “0.450”. Therefore, the process ends.

As described above, the speech model learning device according to the present embodiment performs discrimination learning for a pair of single syllables. As described above, in the present embodiment, the learning dictionary is narrowed down to the minimum two syllables by using the property of discriminative learning in which the accuracy of the model depends on the learning dictionary.
A high-accuracy syllable identification model that specializes only in syllable identification can be generated. Therefore, if such discriminative learning is performed for each pair of single syllables for all supposed voice pairs, the learning result (voice model) is combined with any word in the specific language. Voice recognition becomes possible.

Next, a description will be given of a speech model learning apparatus according to a second embodiment of the present invention. The speech model learning device according to the present embodiment is basically the same as the speech model learning device according to the first embodiment. However, the contents stored in the dictionary unit 102 are different from those in the first embodiment. That is, in the present embodiment, since a predetermined syllable string is used as the learning utterance, the dictionary unit 102
Holds a syllable string model representing the content of the input speech as a recognition dictionary. The dictionary unit 102 holds a model that accepts an arbitrary syllable string (hereinafter, a general-purpose model).

The general-purpose model used here is completely different in form from a normal word model. More specifically, a normal word model is represented as a straight chain of subword (syllable) models, whereas a general-purpose model is represented by a network shape as shown in FIG. That is, the general-purpose model has a path from the input side to the output side and a path (null) from the output side to the input side corresponding to all syllables (50 Japanese syllables). The speech model learning device searches for a path that best matches the input speech for such a general-purpose model. For example, when “kettle” is given as the input voice, the voice model learning device searches for the path shown in FIG. In other words, "ya" → "null" → "ka" →
Find the path "null" → "n". The found path can be regarded as a straight chain like a normal word model, and the speech model learning device performs identification learning using this as a word model (error model) of a misrecognized word. As described above, since the general-purpose model has a path corresponding to all syllables and a path returning to the input side, it can correspond to any input voice.

In practice, in the word model, one syllable is composed of a plurality of subwords. For example, when one syllable is composed of three subwords, a normal word model is represented as shown in FIG. Therefore, the corresponding general-purpose model is represented as shown in FIG.

In order to recognize the above-mentioned general-purpose model, see "Basics of Speech Recognition" by Kiyohiro Kano (May 1, 1996).
7th), a continuous word recognition algorithm such as the One Pass DP method described on pages 42 to 43 may be used.

Using such a speech model learning device,
When the speech recognition learning is performed, a syllable string model that is recognized by the device as being closest to the training utterance is obtained as a recognition result for the general-purpose model. That is, discrimination learning between the model representing the correct utterance content and the model closest to the actual utterance (that is, a character string that is easily confused) is automatically performed. As described above, in the present embodiment, a correct model and a confusion model are automatically generated, and discrimination learning for a model that is easily confused is performed. In addition, instead of the general-purpose model, a method using a dictionary including all the words that may be confused may be considered, but in this case, the number of words included in the dictionary may be extremely large, and Not.

Further, the speech model learning apparatus can simultaneously learn a plurality of input speeches. In this case, the dictionary unit 102 holds a syllable string model corresponding to each input voice and a general-purpose model corresponding to each of them. Then, these general-purpose models become the confusion models closest to each input voice by executing the voice recognition learning.

In the present embodiment, it is possible to generate dedicated models of various categories depending on how to construct a general-purpose model. For example, if the general-purpose model is a model representing an arbitrary single syllable, and learning is performed using a single syllable voice, an identification model dedicated to a single syllable can be created. Further, if the general-purpose model is a model representing an arbitrary continuous numeral, an identification model dedicated to the continuous numeral can be created. As described above, according to the present embodiment, it is possible to generate a speech-direction identification model of a certain specific category only by two models of identification learning.

Next, a speech recognition apparatus according to a third embodiment of the present invention will be described with reference to FIGS. As shown in FIG. 8, the speech recognition apparatus according to the present embodiment receives speech as input, and calculates and outputs acoustic features for each fixed time (frame), and outputs the result.
And a dictionary unit 502 for holding a pair of word model dictionaries D for speech recognition, and a feature amount for each frame output from the speech analysis unit 501, and each word model of the dictionary unit 502 based on the feature amount. And an input voice, a calculation unit 503 for calculating and outputting the likelihood, and an output unit 504 for obtaining a recognition result from the likelihood output from the calculation unit 503 and outputting the result.

Here, as shown in FIG. 9, for example, the dictionary unit 502 holds two word models to be recognized and subword models constituting each of the word models. In the example of FIG. 9, a word model is held for each of the words “yoshida” and “uchida”. Each word model is configured by connecting syllable models constituting each word in time series. That is, "Yoshida"
A model M1 representing “yo”, a model M2 representing “shi”, and a model M3 representing “da”.
It is composed of a model M4 representing "U", a model M5 representing "chi", and a model M6 (= M3) representing "da". Here, these models M1 to M6 are speech models obtained by the speech model learning device according to the first embodiment. That is, the models M1 and M4 are speech models obtained by discriminative learning using a pair of the syllable “yo” and the syllable “u”. The models M2 and M5 are speech models obtained by discriminative learning using a pair of syllable "shi" and syllable "chi". However, for models M3 and M6,
Since it cannot be obtained with the speech model learning device according to the first embodiment, it is necessary to prepare it separately using, for example, an EM algorithm.

Each of the single syllable models M1 to M6 is represented as a time series of a subword model. Here, an example composed of 9 subwords is shown. Each subword includes information representing a feature amount for each frame and its standard deviation.

Hereinafter, the operation of the speech recognition apparatus according to the present embodiment will be described with reference to FIG.

First, the voice analysis unit 501 determines in step S7
In step 1, the input voice is divided into frames, the frequency of the voice in the section is analyzed for each frame, and the feature amount is calculated and output. As the feature amount, audio power, power change amount, cepstrum, cepstrum change amount, and the like can be used.

Next, the calculating unit 503 determines in step S72
The dictionary unit 5 uses the input feature amount of each frame.
The likelihood is calculated for each word model in the dictionary stored in the second dictionary and output to the output unit 504. Here, as a method of obtaining likelihood, dynamic programming (DP)
Dynamic time warping:
DTW) is available. For this DTW, for example,
Kiyohiro Kano, "Basics of Speech Recognition" (May 17, 1996
Sun), pp. 32-34.

Finally, the output unit 504 determines in step S73
Then, the likelihood of each word model output from the calculation unit 503 is compared, and a word model with a high likelihood is output as a word closest to the input speech (as a recognition result).

As described above, in the present embodiment, word recognition is performed using the speech model obtained by the speech model learning apparatus according to the first embodiment. And high-accuracy speech recognition can be realized without performing In other words, speech recognition can be performed with high accuracy without performing recognition learning on an arbitrary word.

Next, a speech recognition apparatus according to a fourth embodiment of the present invention will be described. Although the speech recognition device according to the third embodiment is a device that recognizes a pair of words held by the dictionary unit 502, the speech recognition device according to the present embodiment can recognize three or more words. .

In this embodiment, the dictionary unit 502
Holds a dictionary D including three or more word models.
Here, each word model is a model obtained by performing syllable pair recognition learning using the speech model learning apparatus according to the first embodiment, and is connected in time series. When three pairs of word models handle four words, six pairs of word models are held as the dictionary D.

The calculation unit 503 calculates the likelihood for each word for all the word pairs included in the dictionary D, as in the third embodiment. At this time, the word pair (s,
t), the likelihood of each word is L
Assuming that (s) and L (t), the relative likelihood of the word s with respect to the word t with respect to the word pair (s, t) is expressed by Expression 6.

[0064]

(Equation 6) As a result, the relative likelihood average of the word s is given by Expression 7.

[0065]

(Equation 7) The calculation unit 503 calculates the relative likelihood average for each word, and outputs the average to the output unit 504.

The output unit 104 compares the relative likelihood averages output from the calculation unit 503 and outputs a word model having the largest value as a recognition result.

Thus, the speech recognition apparatus according to the present embodiment can recognize three or more words.

[0068]

According to the present invention, the learning dictionary is narrowed down to a minimum of two syllables by utilizing the property of discriminative learning in which the accuracy of the model depends on the learning dictionary. A highly accurate syllable identification model can be generated. Therefore, assuming a specific language as a target of speech recognition, if a speech pair necessary for speech recognition is specified, and a speech model is generated for all specified syllable pairs,
A speech recognition device having high accuracy for any word can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a flowchart illustrating an operation of the speech model learning device in FIG. 1;

FIG. 3 is a diagram for explaining a single syllable model and a subword model.

FIG. 4 is a diagram for explaining an operation of a control unit in FIG. 1;

FIG. 5 is a diagram for explaining a general-purpose model.

FIG. 6 is a diagram for explaining selection of a path for a general-purpose model.

FIG. 7A is a diagram for explaining an actual word model, and FIG. 7B is a diagram for explaining an actual general-purpose model.

FIG. 8 is a block diagram showing a third embodiment of the present invention.

9 is a diagram illustrating a dictionary D of the dictionary unit 502 in FIG.

FIG. 10 is a flowchart for explaining the operation of the speech recognition device in FIG. 8;

[Explanation of symbols]

 Reference Signs List 101 voice analysis unit 102 dictionary unit 103 pattern matching unit 104 calculation unit 105 correction unit 106 control unit 501 voice analysis unit 502 dictionary unit 503 calculation unit 504 output unit

Claims (9)

[Claims] A speech analysis unit for calculating a feature amount of the input speech at predetermined time intervals; a dictionary unit storing a plurality of learning word patterns; Match each word pattern,
A pattern matching unit that calculates the association and likelihood with each learning word pattern; a calculation unit that calculates the parameter correction amount of each learning word pattern based on the association and the likelihood; A correction unit for respectively correcting the parameters of the plurality of word patterns stored in the dictionary unit, wherein the dictionary unit stores a pair of different monosyllable models as the learning word pattern. Characteristic speech model learning device. 2. Based on the likelihood, it is determined whether or not to perform a learning process again using a learning word pattern having a parameter corrected by the correcting unit. The speech model learning device according to claim 1, further comprising a control unit that instructs the matching unit to execute the learning process again. 3. The dictionary unit holds a pair of word models as the word patterns instead of the pair of different monosyllable models, and one of the pair of word models corresponds to a predetermined input voice. 3. The speech model learning device according to claim 1, wherein the word model is a word model, and the other word model is a word model closest to the predetermined input speech obtained by a continuous word recognition algorithm. 4. The dictionary unit holds a plurality of pairs of word models as the word patterns in place of the pair of different monosyllable models, and one of the paired word models has a different input model. 3. The speech model according to claim 1, wherein the other word model of each pair is a word model closest to the predetermined input speech obtained by a continuous word recognition algorithm. 4. Learning device. 5. A voice analysis unit for calculating a feature amount of input speech at predetermined time intervals, a dictionary unit for storing a pair of word patterns, the feature amount at predetermined time intervals, and the pair of word patterns. A calculating unit that performs matching with each of the above, and calculates the likelihood of each of the pair of word patterns,
An output unit that determines and outputs one of the pair of word patterns corresponding to the input voice based on the likelihood, and the pair of word patterns stored in the dictionary unit each include a monosyllable model. It is configured by connecting in the order of utterance,
A speech recognition apparatus, wherein each single syllable model is generated by discriminative learning using a pair of learning word patterns paired with a single syllable model of the other word pattern in the same utterance order. 6. A speech analysis unit for calculating a feature amount of the input speech at regular time intervals, a dictionary unit for storing word patterns corresponding to a plurality of words, the feature amount at regular time intervals, A matching unit that performs matching with each of the word patterns and calculates the likelihood of each word pattern, and an output unit that obtains and outputs the word pattern closer to the input voice from the likelihood of the word pattern, The word patterns corresponding to the plurality of words are formed by connecting monosyllable models constituting each word in the order of utterance, and each of the monosyllable models constituting each word constitutes another word Each word has a number of word patterns corresponding to the number of other words, so that each word is generated by discriminative learning using a pair of learning word patterns that form a pair in each utterance order. Prepared A speech recognition device characterized by the following. 7. A first step of calculating a feature amount for each fixed time with respect to an input single syllable utterance, and a step of calculating a pair of learning single syllable patterns stored in advance with the feature amount for each fixed time. A second step of performing matching with each of the syllable utterances and each learning single syllable pattern and obtaining a likelihood, and a parameter of each learning single syllable pattern based on the correspondence and the likelihood. A third step of calculating a correction amount; and a fourth step of respectively correcting parameters of the plurality of learning single syllable patterns stored in advance based on the parameter correction amount. Voice model learning method. 8. After the fourth step, it is determined based on the likelihood whether or not the learning process is to be performed again using the learning word pattern having the corrected parameter, and the learning process is performed again. 8. The speech model learning method according to claim 7, further comprising a fifth step of repeating the second step and the subsequent steps when performing. 9. A first step of calculating a feature amount of the input voice at regular time intervals, and performing matching between the feature amount at regular time intervals and each of a pair of word patterns stored in advance. A second step of calculating the likelihood of each of the pair of word patterns; and a third step of determining and outputting one of the pair of word patterns as corresponding to the input voice based on the likelihood. 8. A speech recognition method comprising using a word model to which a single syllable model obtained by the speech model learning method according to claim 7 is connected as the pair of word patterns.