JP3231365B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus capable of effectively improving the recognition performance for word speech, sentence speech, and the like.

[0002]

2. Description of the Related Art Conventionally, in a speech recognition apparatus for a word or a sentence, phonemes, syllables, vowels / consonants have been used because of the easiness of changing the recognition vocabulary and the easy use of syntax and semantic information.・ Recognition methods for recognizing speech in units of vowels and the like have been studied.

[0003] The phoneme referred to here is a linguistic minimum unit in voice utterance, and is a vowel / I in Japanese.
/, / E /, / A /, / O /, / U /, sound repellent / N /, nasal consonants / m /, / n /, plosive consonants / P /, / T /, / K
/, / B /, / D /, / G /, semi-vowel / Y /, / W /, etc., and are also called phonemes.

A syllable means a consonant / vowel, a consonant / half vowel / vowel, and the like. Here, a description will be given of a phoneme as an example, but the same applies to a case where units such as syllables, vowels / consonants / vowels are used.

[0005] Recognition of input speech is first performed on phonemes, and then recognition of words and sentences is performed. In phonological recognition, bottom-up processing that performs temporal segmentation using each feature for each phoneme, and generation of words and sentences without using segmentation using higher-level knowledge such as recognized vocabulary, syntax, and meaning. There is a top-down process that simultaneously performs segmentation and recognition.

However, it is difficult to classify phonemes in advance with high accuracy, and therefore, top-down processing is often used. In order to achieve a high recognition rate by top-down processing, it is important to improve the phoneme recognition rate, and many methods have been used.

[0007] For example, in the TDNN (Time Delayed Neural Network) method using a neural network, phoneme data (hereinafter, referred to as labeled phoneme data) that has been segmented in advance is used for the neural network. Recognition and erroneously recognized data are repeatedly subjected to learning to change the network coefficients so as to be correctly recognized, thereby improving the phoneme recognition rate. However, in this method, phoneme recognition is performed using only pre-segmented phoneme data, so that top-down phoneme recognition is not always performed with high accuracy. This is because the segmentation of labeled phoneme data and the segmentation of top-down word and sentence recognition differ in phoneme error correction learning. In addition, when a dictionary for phoneme recognition is created based on data previously segmented by an expert as labeling phoneme data, segmentation of labeling phoneme data and segmentation of top-down recognition can be performed. different.

Therefore, attempts have been made to improve the phoneme recognition rate by using a method of creating a phoneme dictionary by performing phoneme segmentation again based on the results of top-down recognition. Therefore, a high phoneme recognition rate and a high recognition rate of the input speech have not been obtained because learning of a phoneme dictionary based on the speech cannot be performed.

As described above, in the prior art, a device is developed without sufficiently considering a phoneme dictionary learning for improving phoneme recognition and temporal division in top-down recognition. Was difficult to obtain. Further, in a speaker-adaptive device that adapts to a speaker using a voice registered by a user, adaptive learning can be performed so that the speaker can be adapted by a phonological dictionary learning method. As a result of the development of the apparatus without sufficiently considering such temporal segmentation, the effect of speaker adaptation is not sufficient, and therefore, the performance of the recognition apparatus is insufficient.

Further, in the apparatus for recognizing speech in noise, since the feature vector of the phoneme changes due to the noise and the time division of the phoneme feature changes, the recognition apparatus taking into account the temporal segmentation as described above. Was undeveloped. For this reason, practical vocabulary recognition devices, continuous speech recognition devices, and speech word processors based on phonemes could not be put to practical use or expanded in application.

[0011]

Speech recognition is first performed for phonemes, and then words and sentences are recognized. However, it is difficult to improve the accuracy of phoneme recognition. This is because conventional phoneme-based speech recognition systems did not sufficiently consider the time segmentation of input speech into phonemes in top-down recognition using phonological dictionary learning to improve the phoneme recognition rate and top-level knowledge. As a result, a phonemic recognition dictionary cannot be created with high accuracy, and a high recognition rate cannot be obtained in a phonemic-based speech recognition device.

That is, top-down phoneme recognition is performed using a phoneme dictionary, phoneme segmentation is performed again based on the result, a phoneme dictionary is created, and phoneme recognition is performed again based on the dictionary. Therefore, since the learning of the phoneme dictionary based on the phoneme recognition result cannot be performed, a high phoneme recognition rate and a high recognition rate of the input speech have not been obtained.

The present invention has been made in consideration of such circumstances, and has as its object to consider both phonological dictionary learning and temporal segmentation of input speech into phonology in top-down recognition. It is an object of the present invention to provide a speech recognition device capable of creating a high-performance phonological dictionary and ensuring a high recognition rate of input speech.

[0014]

In order to achieve the above object, the present invention is configured as follows. That is, a phoneme feature vector is continuously extracted by a phoneme feature vector acquisition means from a time series of speech feature parameters obtained by dividing input speech data, and a phoneme feature vector obtained thereby and a phoneme dictionary storage means are stored. The likelihood time series of the matching phoneme is obtained by matching the phoneme dictionary composed of the phoneme feature vectors for each phoneme to be obtained by the phoneme matching means, and the obtained likelihood time series of the matching phoneme and the word dictionary stored in the word dictionary storage means. And the likelihood of the word is determined by the word matching means, and the word having the best likelihood is obtained as a recognition word.The phoneme dictionary storage means enables the phoneme dictionary to be updated. A learning voice data holding unit for holding learning voice data used in a word learning mode; and a learning voice data holding unit as the input voice data in a word learning mode. A phoneme feature vector selection unit for selecting a phoneme feature vector of the learning speech data from the input phoneme segment information obtained by the word matching means for the section using voice data, and obtaining a phoneme feature vector for training; A learning phoneme feature vector storage unit for storing the learned learning phoneme feature vector, and the learning phoneme feature vector and the learning phoneme feature vector of the learning phoneme feature vector storage unit in the phoneme learning mode. A phoneme matching determination unit that compares the result of matching with the phoneme dictionary, a phoneme unit dictionary updating unit that updates a phoneme dictionary in the phoneme dictionary storage unit based on the matching result from the phoneme matching determination unit, A learning control unit that controls to switch between a learning mode and the phoneme learning mode.

Alternatively, the present invention is characterized in that the learning control unit has a control function of repeatedly executing the word learning mode and the phonological learning mode using the updated phonological dictionary and updating the phonological dictionary. It is assumed that.
Alternatively, the present invention is characterized in that the learning control unit has a control function of controlling the number of repetitions of the word learning mode and the phoneme learning mode.

[0016]

According to the present invention, a phoneme feature vector is continuously extracted from a time series of speech feature parameters obtained by dividing input speech data, and the phoneme feature vector is compared with a phoneme dictionary. Then, a phoneme likelihood time series is obtained, and the obtained phoneme likelihood time series is collated with the word dictionary.
Along with outputting the recognition result, the input phoneme classification information is obtained from the matching result between the phoneme likelihood time series and the word dictionary,
A learning phoneme feature vector is selected from the time series of phoneme classification information and input feature parameters, and the learning phoneme feature vector is compared with the phoneme dictionary, and the phoneme dictionary is learned from the phoneme matching determination result and the training phoneme feature vector. To get a new phonemic dictionary.

[0017]

[0018]

The learning mode includes a "phonological learning mode" and a "word learning mode". The "word learning mode" is a mode for selecting a learning phonological feature vector from the result of word matching. The "mode" is a mode for learning a phonemic dictionary using the stored phonemic feature vectors for learning. In the "word learning mode", a time series of a time series of input feature parameters is obtained from a result of collation with a word dictionary expressing the content of input speech, and a feature vector corresponding to each phoneme in the input word dictionary is obtained. In the "phonological learning mode", the phonological learning mode is compared with the phonological feature vector for learning, and it is determined whether or not the data is necessary for learning. Used to learn the phonetic dictionary, got a new phonemic dictionary,
Learning control is performed by repeating learning in the phonological learning mode, and if necessary, selecting a learning phonological feature vector in the "word learning mode" from the learning result, and further performing learning in the "phonological learning mode". The process is repeated, and if the learning is sufficiently performed, the learning is stopped.

The word dictionary referred to here is for performing word recognition. For example, when performing sentence recognition, a sentence recognition dictionary generated from vocabulary, syntax, semantic information and the like is used. Similarly, the “word learning mode” indicates a mode in which word recognition is performed. For example, in the case of performing sentence recognition, it indicates the “sentence learning mode”.

As described above, according to the present invention, the time series of the time series of the characteristic parameters of the input is obtained from the result of collation with the word dictionary expressing the contents of the input speech, and the time series corresponding to each phoneme in the input word dictionary is obtained. Select and memorize the feature vector, and further compare the learning phoneme feature vector with the phoneme dictionary,
It is determined whether or not the data is necessary for learning, and if the data is necessary for learning, the data is used for learning a phonological dictionary, so that a new phonological dictionary is obtained.
High-performance phonemic collation can be performed in speech recognition in units of phonemes.

In particular, in the "word learning mode", a temporal division suitable for a time-series phonological dictionary of the input characteristic parameters is obtained from the result of collation with the word dictionary, and a characteristic vector corresponding to each phonological element is extracted with high accuracy. In the phonological learning mode, phonological dictionary learning can be performed using the phonological feature vector for learning and the phonological dictionary, and a phonological dictionary with high performance can be created. Since the configuration is realized by sequentially repeating the learning combination of the “word learning mode”, the extraction of the phoneme feature vector and the creation of the phoneme dictionary can be performed simultaneously with high accuracy.

Further, in the speaker-adaptive recognition device, by repeatedly performing phonological learning and word learning, it is possible to create a phonological dictionary adapted to the speaker while taking into account the characteristics of the pronunciation method of the speaker. High recognition performance can be obtained.

Further, in the apparatus for recognizing speech in noise, it is possible to create a phonemic dictionary in consideration of changes in the phonetic feature vector and changes in the temporal division of phonemic features due to noise, and high recognition performance can be obtained. .

Therefore, according to the present invention, it is possible to create a high-performance phonological dictionary in consideration of both phonological dictionary learning and temporal segmentation of input voice into phonology in top-down recognition, and high recognition of input voice. It is possible to provide a speech recognition device capable of securing a rate.

[0026]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition apparatus according to one embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a basic schematic block diagram of an embodiment apparatus for performing word recognition. In the figure, 1 is a voice input / fraction unit, 2 is a phoneme feature vector extraction unit, 3 is a phoneme matching unit, 4 is a word matching unit, 5 is a phoneme dictionary unit, 6 is a word dictionary unit, and 7 is a phoneme dictionary updating unit. , 8 is a phoneme matching determination unit, 9 is a learning phoneme feature vector storage unit, 10 is a learning voice data storage unit, 11 is a learning control unit, 12 is a matching result determination output unit, and 13 is a phoneme feature vector selection unit. .

SW1 to SW5 denote signal path switching function units, respectively.
Among the paths selected in the above, the path indicated by the symbol A is the path in the “recognition mode”, and the path indicated by the symbol B is the path in the “phonological learning mode”. ,
Also, the path indicated by the symbol C indicates a path in the “word learning mode”. Such path switching control is controlled by the learning control unit 11.

The voice input / fraction unit 1 includes a low-pass filter (LPF) for removing high-frequency components of 5.4 KHz or more from voice data, and a voice signal input via the LPF at a sampling frequency of 12 KHz and quantization. Bit number 16bi
A / D converter for converting a digital signal at t
The digital signal output by the / D converter is divided into 256 points by DFT (Discrete Fourier Transform) to obtain a filter bank output of 16 channels obtained by dividing the frequency into 16 from the frequency spectrum of 128 points every 8 msec. A demodulation unit that performs logarithmic processing and converts it to time-series data of 16-dimensional speech feature parameters, and a speech signal directly input from a microphone or the like, or a learning speech data storage unit. The learning voice data stored in the memory 10 is received as a voice input, and is converted into time-series data of voice characteristic parameters and output.

The phoneme feature vector extraction unit 2 receives the time-series data of the feature parameters output from the speech input / fractionation unit 1 and stores the time-series data of the feature parameters in the phoneme dictionary unit 5. The phoneme feature vector for matching with the phoneme dictionary is sequentially extracted and given to the phoneme matching unit 3 or the phoneme feature vector selection unit 13. The phoneme matching unit 3 uses a well-known matching method to obtain the phoneme feature. The vector is collated with the phoneme dictionary of the phoneme dictionary unit 5 to output a likelihood time series. The likelihood time-series output is input to the phoneme matching determination unit 8 in the “phoneme learning mode” (B), and is input to the word matching unit 4 in the “recognition mode” (A).

The phoneme dictionary unit 5 stores a phoneme dictionary having various phoneme information, and the phoneme dictionary updating unit 7 can update the phoneme dictionary.

The phoneme collation judging unit 8 compares the collation result from the phoneme collation unit 3 with the phoneme name assigned to the learning phoneme feature vector in the learning phoneme feature vector storage unit 9 and determines whether or not both match. Is output to the learning control unit 11, and the determination result as to whether or not to update the phonemic dictionary based on whether or not the predetermined matching condition is satisfied is sent to the phonemic dictionary updating unit 7. It has a function of outputting.

The phoneme dictionary updating unit 7 has a function of selecting a feature vector used for phoneme dictionary learning from the judgment result from the phoneme feature vector storage unit 9 for learning, providing the selected feature vector to the phoneme dictionary unit 5, and updating and storing the same.

The word matching unit 4 receives the likelihood time-series output including the phoneme name from the phoneme matching unit 3, matches the likelihood with the word dictionary of the word dictionary unit 6, and obtains the phoneme classification information (phoneme segmentation position) and the phoneme name. And the likelihood of phonemes, likelihood as words, and information such as the total score.
Receives the result of the collation by the word collating unit 4 and outputs the word having the best score as a recognition result.

The learning control unit 11 controls the processing of the "word learning mode" and the "phonological learning mode", determines the end of the phonological learning, and controls the learning while controlling the data flow in response to the instruction to start learning. The end is determined, and learning control is performed.

In the word learning mode, the phoneme feature vector selection unit 13 uses the phoneme classification information provided by the word matching unit 4 and the phoneme name, and uses the phoneme feature of the training speech data extracted by the phoneme feature vector extraction unit 2. It has a function of classifying the vector, storing the classified phoneme feature vector in the learning phoneme feature vector storage unit 9 with the phoneme name given thereto.

The learning voice data storage unit 10 stores learning voice data, and includes voices for various reference words, readings (time-series information of phoneme names) for the words, and phoneme classification information. Are stored together with the supplementary information.

Next, the operation of the present apparatus having such a configuration will be described with reference to FIG.
This will be described with reference to the flowchart of FIG. In this system, there are two modes, "recognition mode" and "learning", and "learning" is composed of two types of "word learning mode" and "phonological learning mode". The “recognition mode” is a mode for performing a process of recognizing an input voice, is a process including steps S22 to S24, and is a normal execution mode. “Learning” is a learning mode for improving the accuracy of the phonemic dictionary, and is realized by a pair of “word learning mode” and “phonological learning mode”.

The "word learning mode" is a mode for knowing how much words can be recognized in the current phoneme dictionary, and is a process including steps S31 to S33. A mode in which phonemes are re-learned based on the word recognition results in the "word learning mode",
This is a process including steps S34 to S38.

The voice is first input to the voice input / separation unit 1 in the form of a voice signal from voice input means such as a microphone. The voice input / fraction unit 1 removes high-frequency components from the input signal, converts it into digital data, and obtains a time series of 16-dimensional voice feature parameters using the digital data.

That is, the voice input / separation unit 1 includes a low-pass filter (LPF) for removing high-frequency components of 5.4 kHz or more from voice data, and a voice signal input via the LPF at a sampling frequency of 12 kHz and a quantum 16 bits
an A / D converter for converting a digital signal in bits,
By DFT (Discrete Fourier Transform) analysis of 56 points,
From the frequency spectrum of 128 points every msec, 16
And a dividing unit that obtains a filter bank output of 16 divided channels and obtains a time series of 16-dimensional voice feature parameters by logarithmic processing. Output in the form of a time series.

The time series of the feature parameters of the speech output from the speech input / fractionation unit 1 in this manner is input to the phoneme feature vector extraction unit 2 (S21).

On the other hand, the learning control unit 11 first selects a mode according to an instruction as to which of the three modes to use, and sets the operation mode of the system. This instruction can be set, for example, by operating a setting means (not shown) by an operator or a user of the system, and automatically enters the learning mode when the error rate of the recognition result reaches a predetermined value. Or the recognition mode is executed while learning is performed.

When the "recognition mode" is instructed, the learning control unit 11 sets the switches SW1, SW3, SW4, and SW5 so as to be connected by the path indicated by the symbol "A". Thereby, the path of the speech input / analysis unit 1, phoneme feature vector extraction unit 2, phoneme matching unit 3, word matching unit 4, and matching result determination output unit 12 becomes operable.

Therefore, in the "recognition mode", which is a mode for recognizing an arbitrary voice input from the outside via a microphone or the like, the voice input / separation unit 1 is used as described above.
The phoneme feature vector extraction unit 2 receives the time series of the feature parameters of the speech output from the phoneme and outputs the phoneme feature vectors for collation with the phoneme dictionary stored in the phoneme dictionary unit 5 from the time series of the feature parameters. Extract and phoneme matching unit 3
Send to

The phoneme matching unit 3 compares each of the phoneme feature vectors sequentially inputted with the phoneme dictionary stored in the phoneme dictionary unit 5, and outputs a likelihood time series of phonemes recognized as corresponding (S22). ). Conventionally, the matching method is
Various proposed methods can be appropriately adopted, and for example, can be obtained by a Mahalanobis distance, a composite similarity method, or the like.

Recognized phonemes by various sections obtained by the collation in the phoneme collation unit 3 and information such as their likelihood and phoneme division information are sequentially output to the word collation unit 4, and the word collation unit 4 uses the original information. In addition to performing word matching using a word dictionary, the likelihood, total score, likelihood of phonemes, and the like are calculated for each word (S23). The likelihood referred to here is similarity,
Distances, probabilities, and the like are converted by various methods, and are determined by a collation method.

The result of word collation by the word collating unit 4 is determined by the collation result determination output unit 12, and a word showing the best overall value is output as a recognition result (S24).

Here, the configuration of the phoneme feature vector is a fixed-dimensional vector. For example, a time series of feature parameters of 16 channels of audio is continuously arranged in a time axis direction, for example,
A 16 × 6 = 96-dimensional time-frequency vector using six frames is used. For the phoneme dictionary stored in the phoneme dictionary unit 5, for example, in the matching by the composite similarity method, a correlation matrix of each phoneme feature vector for creating a dictionary is created.
The correlation matrix is composed of eigenvectors and eigenvalues obtained by KL expansion.

The word dictionary stored in the word dictionary unit 6 has a configuration in which the connection of phonemes of words to be recognized is described by a graph having phonemes as nodes. FIG. 3 shows an example of the word dictionary. FIG. 3 is a description example of a dictionary of words of proper nouns “Akita (Akita)”. In the figure, “I” enclosed in a circle indicates “I” which is voiced, and “I” enclosed in a circle. “#I” indicates “I” which is a vowel that has been devoiced.

That is, the recognition word “Akita” shown as an example is a phoneme / A /, / K /, / I /, /
T /, / A /, but in normal pronunciation / K /, / I / (that is, "I"), / I /
Since the voice is between the unvoiced plosives / K / and / T /, the vocal cords are sounded without being vibrated.

This unvoiced vowel has many acoustically high frequency components, and its frequency characteristics are different from those of [I] produced by vibrating (voiced) vocal cords. On the other hand, when carefully uttered, / K /, / I / of / I / is vocalized.

In such an utterance, a phonological sequence to be recognized is represented in a graph based on the characteristic (sound form rule) that the acoustic feature changes. Further, information on each voice segment, for example, information on the minimum duration (1Nn min) and the maximum duration (1Nn max) of the voice segment Nn are also stored in the word dictionary of the word dictionary unit 6. And is used in the word matching unit 4.

Next, the operation of the word collating unit 4 will be described with reference to FIGS. The word matching unit 4 obtains the maximum likelihood value LFWk of each recognition target word Wk from the sum of the likelihoods of the respective phonemes according to the sequence of the phonemes for each recognition target word Wk.

That is, taking the collation of the word Wk as an example,
The phoneme feature vector in the input speech of the word Wk is searched for a phoneme that seems to be applicable while changing the range within the respective phoneme collation ranges, and phoneme classification information and a phoneme name are output (phoneme collation unit). Receiving the classification information and the phoneme name, the likelihood of the phoneme and an overall score based on a predetermined calculation formula are obtained, and the one showing the maximum likelihood is sequentially obtained as the corresponding phoneme from among them, thereby obtaining the maximum likelihood. The maximum likelihood LFWk of the phoneme sequence indicating
Next, the word matching unit 4 outputs a word corresponding to the phoneme string of the obtained maximum likelihood LFWk as a matching result. However,
Each phoneme, for example, the n-th word n (n is 1, 2, 3, 4
...) Phoneme Nn (step S1) duration lNn
Min to lNn have predetermined values for each phoneme (step S2), and represent the duration of the phoneme Nn in normal speech.

As described above, the boundary positions t and t 'between the phonemes are determined by using the dynamic programming method while utilizing the restriction of the duration of the phonemes (min of min. LNn to max of max. LNn, step S3). t 'is the end position of the phoneme section immediately before the current search phoneme section, and t is the end position of the end of the current search phoneme section, while determining the sum from the beginning of the word Wk to the current search phoneme position. Is obtained (LFNn, t = LF)
Nn-1, t '+ PLFNn (t, t') where LFNn, t is LFWk, and the total likelihood from the voice start to t, LFNn-1, t '
Is the cumulative likelihood from the beginning of the voice to t ', PLNn (t, t') is the sum of the log likelihood in the current phoneme section), the likelihood LFWk indicating the maximum value among the likelihoods, and each phoneme in that case. Boundary position t between
And t 'are obtained (step S4).

Then, all the phonemes constituting the recognition target word Wk are collated, and the likelihood LFNn, TE in the phoneme section of the speech end TE is added to obtain the total value of the likelihood from the speech start end to the speech end TE. The likelihood LFWk for the word Wk is set (step S6). The one with the highest likelihood LFWk for the word Wk is used, the phoneme sequence in that case is collated with each word in the word dictionary, and the similarity of each word in the word dictionary to the word Wk is obtained in the form of a total score. The matching result determination output unit 12 determines the word having the highest similarity, and outputs the determined word as a recognition result.

[0058]

By following the processing of this flowchart, in the example of FIG. 4, / A /, / I /, /
For each phoneme of K /, / T /, it is recognized that the sum of the likelihoods takes the maximum value. In this example, /
A /, / I /, / K /, / T /, and / A / were found to have the highest likelihood phonemes, and the boundary positions of the respective phonemes at that time, That is,
Temporal information Ki (i = 1, 2, 3, 3) as shown in FIG.
..) Can be obtained, and this is used as phoneme section information in the word learning mode.

In the learning, as shown in FIG. 2, a "word learning mode" is first executed, and then a "phonological learning mode" is executed.

In the "word learning mode", SW1 to SW5
Is controlled so that the process flows following the path “C”. Therefore, by this, the learning speech data storage unit 10, speech input / analysis unit 1, phoneme feature vector extraction unit 2, phoneme matching unit 3, word matching unit 4, phoneme feature vector selection unit 13, learning phoneme feature vector storage The path of the unit 9 becomes operable.

In the "word learning mode", a database of learning voice data is prepared in advance. The learning voice database includes learning voice data and utterance content information, and is stored in the learning voice data storage unit 10.

When the process enters the “word learning mode”, the learning voice data stored in the learning voice data storage unit 10 is read out and input to the voice input / fraction unit 1.
The speech data for learning is divided by the speech inputting / dividing unit 1 in the same manner as in the "recognition mode", and is converted into time-series data of characteristic parameters of the speech (S21).

Then, the phoneme feature vector extraction unit 2 extracts feature vectors from the time-series data of the feature parameters of the converted speech, and the phoneme matching unit 3 uses the extracted feature vectors to determine the likelihood of the corresponding phoneme. Is obtained (S31). Next, based on this, the word matching unit 4 obtains the corresponding word from the word dictionary.

As described above, a word dictionary is selected from the utterance contents of the learning speech data, and word matching is performed by the same method as described in the "recognition mode" to find a corresponding word (S32).

However, in the "word learning mode", voice data for learning is used, and in the voice data, all information such as phoneme sections, phoneme names, and readings are prepared in advance as accompanying information. Therefore, it is not necessary to perform matching with all the words in the word dictionary stored in the word dictionary unit 6, and it is sufficient to perform matching with the words of the speech selected by learning.

Next, the word matching section 4 outputs the phoneme segment information and the phoneme name by word matching, and the phoneme feature vector selecting section 13 extracts the phoneme feature vector extracting section 2 using the phoneme segment information obtained by the matching. The phoneme feature vector of the obtained learning speech data is selected for the corresponding section, the above-mentioned phoneme name is assigned, and the phoneme feature vector storage unit 9 stores the phoneme feature vector as a training phoneme feature vector (S3).
3).

Various types of information including phoneme classification information obtained by word matching in the word matching unit 4 are also output to the learning control unit 11 and used to determine the end of learning.

That is, since the accompanying information from the learning voice data storage unit 10 is input to the learning control unit 11,
The word matching unit 4 can know the degree of similarity from the phoneme classification information and other information obtained by the word matching, determine whether or not the recognition has been performed correctly, and complete the learning. Can be determined.
After the processing in the “word learning mode” is completed, the process proceeds to the “phonological learning mode”. In the “phonological learning mode”, learning phonetic feature vectors are sequentially read from the learning phonemic feature vector storage unit 9, and the phoneme matching unit 3 performs matching with the phonemic dictionary (S 34).

The phoneme collation judging unit 8 compares the collation result of the phoneme collation unit 3 with the phoneme name assigned to the learning phoneme feature vector in the learning phoneme feature vector storage unit 9 to determine whether they match. The determination is performed, and if necessary, the likelihood for each phoneme is obtained, and this and the determination result are output to the learning control unit 11. The learning control unit 11 includes a learning voice data storage unit 10.
Then, it is determined whether or not phonological learning is to be performed, taking into account information such as the determination result from the phonological collation determining unit 8 based on the accompanying information from the device (S35).

For example, a condition such as "all phonemes were correctly recognized" or a condition "all are correctly recognized and the likelihood difference from the second place is TH1 (threshold" 1 ") or more" is determined. Such conditions can be considered. If this condition is met, the next learning end determination is made without updating.

As a result of the judgment in the phoneme collation judging section 8,
If the phoneme learning is to be performed, the phoneme dictionary updating unit 7 selects a feature vector used for phoneme dictionary learning from the phoneme feature vector storage unit 9 for learning under the control of the learning control unit 11 (S36). As conditions for this selection, for example, a condition such as “misrecognition” or “misrecognition,
Or the likelihood difference from the second place is TH2 or less ". Furthermore, it is conceivable to limit the data to data with high reliability by adding a condition of “data having a likelihood of TH3 (threshold“ 3 ”) or more”.

The selected phoneme feature vector for learning is used by the phoneme dictionary update unit 7 for updating the dictionary (S37). For example, in the case of matching by the composite similarity method, a new correlation matrix is created from a correlation matrix prepared in advance and a phoneme feature vector for learning, and as described in the recognition mode section, the KL expansion is performed. The updated phoneme dictionary is obtained, and the phoneme dictionary for the learning phoneme is collated with the phoneme dictionary. Then, the word is collated and the execution of the “phoneme learning mode” is repeated.

The "learning" including the "phonological learning" described above
Is determined by the learning control unit 11.
This end determination control is performed using the phoneme division information obtained in the “word learning mode”. For example, if the current learning is the first time, the “phonetic division information at the (I−1) -th time and the phoneme division at the I-th time” All the differences in the classification information are the threshold level T
H5 seconds or less. " Also,
A condition that “the process ends when the number of times of learning i becomes 5” is also conceivable.

When such an end determination condition is satisfied, the learning is ended (S38).

As described above, in the present system, the learning control section 11 controls the flow of data as shown in FIG. The phonological learning end determination and the learning end determination are performed to control the learning.

In this embodiment, word recognition is performed using a word dictionary. However, when a sentence recognition dictionary including word information and grammatical information is used, sentence recognition is also possible. Further, the likelihood in the present embodiment is the similarity of the distance, the probability, or a value obtained by converting them, and is determined by the matching method. However, what is described as maximum and minimum in the description of the likelihood should be read as minimum and maximum in the case of distance.

Further, in the recognition of the word voice, it is effective to detect the end of the voice prior to the word matching from the viewpoint of reducing the calculation amount.

Although the description of the operation of the system of the present invention has been made with reference to word recognition as an example, the present invention is not limited to word recognition but can be applied to recognition of continuous words, phrases and sentences. Also, the explanation was given using the phoneme as the linguistic minimum unit of speech utterance. is there. In addition, the present invention can be appropriately modified and implemented without departing from the scope of the invention.

[0080]

[0081]

[0082]

In short, the present system continuously extracts a phoneme feature vector from a time series of speech feature parameters obtained by dividing input speech data, compares the phoneme feature vector with the phoneme dictionary, and compares the phoneme feature vector with the phoneme dictionary. Get the likelihood time series,
The obtained phoneme likelihood time series is collated with the word dictionary, and a recognition result is output. Further, the phoneme likelihood time series is obtained from the collation result between the word dictionary and the input phoneme division information, and the phoneme division information is obtained. A learning phoneme feature vector is selected from the time series of the input feature parameters, the learning phoneme feature vector is collated with the phoneme dictionary, a phoneme dictionary is learned from the determination result of phoneme matching and the phoneme dictionary for learning, and a new phoneme dictionary is obtained. Is obtained.

The learning mode is composed of a “phonological learning mode” and a “word learning mode”. The “word learning mode” is a mode for selecting a learning phonological feature vector from the result of word matching. The "mode" is a mode for learning a phonemic dictionary using the stored phonemic feature vectors for learning. In the "word learning mode", a time series of a time series of input feature parameters is obtained from a result of collation with a word dictionary expressing the content of input speech, and a feature vector corresponding to each phoneme in the input word dictionary is obtained. In the "phonological learning mode", the phonological learning mode is compared with the phonological feature vector for learning, and it is determined whether or not the data is necessary for learning. Used to learn the phonetic dictionary, got a new phonemic dictionary,
Learning control is performed by repeating learning in the phonological learning mode, and if necessary, selecting a learning phonological feature vector in the "word learning mode" from the learning result, and further performing learning in the "phonological learning mode". The process is repeated, and if the learning is sufficiently performed, the learning is stopped.

The word dictionary referred to here is for performing word recognition. For example, when performing sentence recognition, a sentence recognition dictionary generated from vocabulary, syntax, semantic information, and the like is used. Similarly, the "word learning mode" is a mode for performing word recognition, and for example, represents a "sentence learning mode" for performing sentence recognition.

As described above, according to the present invention, the time series of the time series of the characteristic parameters of the input is obtained from the result of collation with the word dictionary expressing the contents of the input speech, and the time series corresponding to each phoneme in the input word dictionary is obtained. Select and memorize the feature vector, and further compare the learning phoneme feature vector with the phoneme dictionary,
It is determined whether or not the data is necessary for learning, and if the data is necessary for learning, the data is used for learning a phonological dictionary, so that a new phonological dictionary is obtained.
High-performance phonemic collation can be performed in speech recognition in units of phonemes.

In particular, in the "word learning mode", a temporal section suitable for a time-series phonological dictionary of input characteristic parameters is obtained from the result of collation with the word dictionary, and a characteristic vector corresponding to each phonological element is extracted with high accuracy. In the phonological learning mode, phonological dictionary learning can be performed using the phonological feature vector for learning and the phonological dictionary, and a phonological dictionary with high performance can be created. Since the configuration is realized by sequentially repeating the learning combination of the “word learning mode”, the extraction of the phoneme feature vector and the creation of the phoneme dictionary can be performed simultaneously with high accuracy.

Further, in the speaker-adaptive recognition device, by repeatedly performing phoneme learning and word learning, it is possible to create a phoneme dictionary adapted to a speaker while taking into account the characteristics of the pronunciation method of the speaker. High recognition performance can be obtained.

Furthermore, in the apparatus for recognizing speech in noise, it is possible to create a phonemic dictionary in consideration of changes in the phonetic feature vector and changes in the temporal division of phonemic features due to noise, and high recognition performance can be obtained. .

Therefore, according to the present invention, it is possible to create a high-performance phonological dictionary taking into account both the phonological dictionary learning and the temporal segmentation of the input voice into phonology in top-down recognition, and high recognition of the input voice. It is possible to provide a speech recognition device capable of securing a rate.

[0091]

As described above, according to the present invention, in speech recognition in units of phonemes, a learning function that takes into account both phoneme dictionary learning and temporal segmentation of input into phonemes by top-down recognition. Thus, high-performance phonemic collation can be performed, and a high recognition rate can be obtained in recognition of input speech using a word dictionary.

[Brief description of the drawings]

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

FIG. 2 is a flowchart for explaining the operation of the system of the present invention.

FIG. 3 is a diagram for explaining a configuration example of a word dictionary used in the system of the present invention.

FIG. 4 is a diagram for explaining the relationship between speech feature parameters and phoneme likelihood in the system of the present invention.

FIG. 5 is a flowchart illustrating a word recognition technique used in the system of the present invention.

FIG. 6 is a flowchart illustrating a word recognition method used in the system of the present invention.

[Explanation of symbols]

1 ... Speech input / fractionation unit 2 ... Phonological feature vector extraction unit
3: Phoneme matching unit, 4: Word matching unit, 5: Phoneme dictionary unit, 6
... Word dictionary unit, 7 phoneme dictionary update unit, 8 phoneme collation judgment unit, 9 phoneme feature vector storage unit for learning, 10 speech data storage unit for learning, 11 learning control unit, 12 collation result judgment output Part 13, phoneme feature vector selection part, SW1
SW5: Signal path switching function unit.

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 15/00-15/28

Claims (3)

(57) [Claims] A phoneme feature vector is continuously extracted from a time series of feature parameters of speech obtained by dividing input speech data by phoneme feature vector acquisition means, and a phoneme feature vector obtained thereby and a phoneme dictionary are stored. A phoneme dictionary composed of phoneme-specific phoneme feature vectors held by the means is compared with a phoneme matching means by a phoneme matching means to obtain a likelihood time series of a matching phoneme,
The likelihood time series of the obtained matching phoneme is compared with the word dictionary held by the word dictionary storage means by the word matching means,
What is claimed is: 1. A speech recognition apparatus comprising: determining a likelihood of a word; and obtaining a word having the highest likelihood as a recognition word, wherein said phoneme dictionary storage means is capable of updating said phoneme dictionary and is used for learning in a word learning mode. Retain audio data
Learning voice data holding means, and the learning as the input voice data in a word learning mode.
Input data obtained from the word matching means using
Phonological features of the speech data for learning from phonological classification information of power
Select a vector for the section,
Phoneme feature vector selection unit for obtaining a vector, and learning for storing the selected phoneme feature vector for learning.
Phoneme feature vector storage unit for learning, and the phoneme feature vector storage unit for learning in phoneme learning mode.
Of the learning phoneme feature vector and the learning phoneme feature vector
Collation in which the vocals match the phonological dictionary in the phonological matching unit
A phoneme collation judging unit for comparing the result with the phoneme dictionary,
A phoneme dictionary update unit for updating a phoneme dictionary in a storage unit, and switching between the word learning mode and the phoneme learning mode.
And a learning control unit for performing control.
Voice recognition device. 2. The learning control section according to claim 1, wherein
The word learning mode and the phonological learning mode using a calligraphy
And a control function for updating the phonological dictionary.
The speech recognition device according to claim 1, further comprising: 3. The learning control section according to claim 1, wherein
And the number of repetitions of the phoneme learning mode
3. A control function according to claim 2, wherein the control function is provided. Voice of
Recognition device.