JPS6053998A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device for automatically recognizing speech signals uttered by a human voice.

Conventional configurations and their problems A speech recognition device that automatically recognizes speech is considered to be very effective as a means for providing data and commands from humans to computers and various machines.

The slam matching method is often adopted as the operating principle of speech recognition devices that have been researched or published in the past.

In this method, standard patterns are memorized in advance for all types of words that need to be recognized, and the degree of matching (hereinafter referred to as similarity) is calculated by comparing them with the unknown input button. , it is determined that the word is the same as the standard baton that yields the maximum match.In this batan matching method, a standard baton must be prepared for every word to be recognized, so the speaker If the button changes, it is necessary to input a new standard button and memorize it. Therefore, if more than several hundred types of words are to be recognized, it will take time and effort to pronounce and register all types of words, and it is expected that the memory capacity required for registration will be enormous. . Furthermore, there is a drawback that the time required for matching the input button and the standard button increases as the number of words increases.

On the other hand, the method of dividing input speech into phoneme units and recognizing them as combinations of phonemes (hereinafter referred to as phoneme recognition) and measuring the similarity with a word dictionary written in phoneme units requires a large amount of memory capacity for the word dictionary. For example, the utterance of ``red'' is /
a/.

AKA by combining the three phonemes /ni/ and /i/
Since it can be expressed in the extremely simple format I, it is easy for unspecified speakers to handle speech with many words.

FIG. 1 shows a block diagram of a speech recognition system characterized by performing phoneme recognition. Audio input through a microphone or the like is analyzed by an acoustic analysis section 1.

As the analysis method, a group of bandpass filters and linear predictive analysis are used to obtain spectrum information for each frame period (about 1 ms). The phoneme discrimination unit 2 uses the spectrum information obtained by the acoustic analysis unit 1 to discriminate phonemes for each frame based on the data in the standard pattern storage unit 3. Standard pattern storage section 3
The standard patterns stored in are prepared in advance for each phoneme of the voices of many speakers. The segmentation unit 4 detects speech sections and determines boundaries for each phoneme (hereinafter referred to as segmentation) based on the analysis output of the acoustic analysis unit 1. The phoneme recognition section 5 performs a task of determining what phoneme is for each phoneme section based on the results of the segmentation section 4 and the phoneme discrimination section 2. As a result, a series of phonemes is completed.

The word recognition unit 6 compares this phoneme sequence with a word dictionary 7 similarly written in phoneme sequences, and outputs the word with the highest degree of similarity as a recognition result.

The most important point when using the above method to target unspecified speakers is to stably obtain high speech recognition accuracy for any speaker environment. Furthermore, this should not place too much burden on the speaker or require expensive components in the speech recognition device.

However, the speech recognition devices that have been announced or prototyped so far have had the drawback of not meeting the above conditions. As a conventional example, a method that targets prediction residuals (Kano, 2003, preferred [used for vowel recognition in conversational speech) Evaluation of targeted LPG distance scale] Journal of the Institute of Electronics and Communication Engineers 80/6°V OL J-63
fl, A5) Te calculates the maximum parameter 1j (j=' + 2 +...+ p ) (p
is the analysis order), and the prediction residual is expressed as follows: where Sj is an autocorrelation coefficient calculated from unknown input speech. This prediction residual Ni is used as a distance measure for each target phoneme, and the phoneme with the minimum Ni is taken as the discrimination result.

However, in this method, the maximum parameter ``mu'', which corresponds to the standard slam of a phoneme, is simply an average value, so even if a learning function is provided to re-create the ``mu'' to suit the user, variations in pronunciation due to articulatory combination may occur. The present invention solves the above-mentioned drawbacks and is capable of dealing with unspecified speakers, as well as being able to deal with differences in speakers, environments, and language. An object of the present invention is to provide a speech recognition device that can stably obtain high speech recognition accuracy without any problems.

Structure of the Invention The present invention has been made to achieve the above-mentioned objective number, and includes an acoustic analysis section that calculates a spectrum or information similar to it (hereinafter referred to as spectral information) from an audio signal divided into each phoneme, and a coefficient storage unit that stores in advance a standard pattern obtained from a standard speech signal consisting of a person; a discrimination filter unit that uses the spectral information and the standard pattern to obtain a filter output for each phoneme; a word dictionary storage unit that stores a written word dictionary; an output unit that compares the degree of similarity or phoneme sequence created through the discrimination filter unit with the word dictionary and outputs the word with the highest degree of similarity as a recognition result; The apparatus includes a learning section that creates a new standard pattern from the result of the output section and the spectrum information of the acoustic analysis section, and rewrites the contents of the coefficient storage section based on the result.

DESCRIPTION OF EMBODIMENTS FIG. 2 shows an embodiment of the configuration of a speech recognition apparatus according to the present invention. The audio signal received from the microphone 31 is sent to the AD converter 21.
Convert to 12 bits with 12 KHz sampling. This is then processed using a signal processing circuit to perform pre-emphasis and 20 ms
A Hamming window is applied, and the linear predictive analysis processor 23 calculates the LPC cepnutrum coefficient every 10 m5. This LPG kepnutrum coefficient is passed through a discrimination filter 24, and the filter output for each phoneme is calculated for each frame and transferred to the main memory 27. The coefficient memory 25 stores filter coefficients for each phoneme.

On the other hand, the bandpass filter 26 calculates band powers and total powers of about three channels, and transfers them to the main memory 27 as data for phoneme segmentation. The main processor 28 uses the results of the discrimination, the filter 24 and the bandpass filter 26 to detect speech intervals and perform segmentation for each phoneme, and then selects the phoneme with the highest degree of similarity from the output of the discrimination filter for each phoneme of the discrimination filter 24. each time, and create a phoneme sequence. By comparing this phoneme sequence with the word dictionary memory 29 which is also written in a phoneme sequence, the word name with the highest degree of similarity is outputted to the output unit 3o as a recognition result.

However, although this alone can be used for unspecified speakers, since the coefficient memory 25 corresponding to the standard pattern is fixed, the recognition performance varies greatly depending on the speaker.
There are cases where the recognition rate becomes considerably low. Therefore, a learning section 32 is provided to provide a new learning function. This learning section 32 receives the LPG cepstral coefficients obtained by the linear predictive analysis processor 23, creates learning data with reference to the results obtained from the output section 30, and creates learning data based on the variance and covariance matrix prepared in advance. An operation is performed to recalculate the discrimination coefficient for each phoneme that is most suitable for the speaker and transfer it to the coefficient memory 25.

Next, the operation of the phoneme recognition device according to the present invention will be explained in detail with reference to FIG.

The vowel /a is converted from a large number of word sounds uttered by multiple speakers inputted in advance from the microphone 31 through the five-word converter 21.
/ +10/ , /u/ , /i/ , /e/ and nasal sounds (
/N/) is cut out.

Using this audio data, the signal processing circuit 22 and linear predictive analysis processor 23 perform linear predictive analysis for each 10 ms analysis interval to calculate p-dimensional LPG cepstral coefficients. Using this LPC kepnutrum coefficient, we can calculate the covariance matrix W for all phonemes and the average value mi for each phoneme.
(L represents the type of phoneme). As a result, the discrimination coefficient aij (1-112,...
..., p) is the inverse matrix 1゛ of the covariance matrix W (jsj')
Letting the element be δjj′, it can be expressed as follows.

alj, mi1', alj, mi' for each phoneme
W' mi (Vt.

) is stored in the coefficient memory 26 as a standard pattern.

Next, a voice whose content is known in advance to the user (for example / a , ' + / 1 / + / u /
1/e/H/o/) is uttered, and the LPO cepnutrum coefficients for each analysis section in the speech section are calculated by the linear prediction analysis processor 23 and transferred to the learning section 32. On the other hand, the degree of similarity is determined by the discrimination filter 24 using the standard pattern stored in the coefficient memory 26 that is stored in advance. The discrimination filter 24 determines the ma...lanobis distance D? with respect to the LPC cepstral coefficients of the input signal. can be expressed as (the first term indicates the transposed matrix), but since the first term is invariant for the phoneme i, the similarity Li can be expressed simply as , and the similarity can be expressed using equation (4). calculate. The result is stored in main memory 2.
7 and creates a phoneme sequence through the main processor 28. Next, a value indicating the position of the phoneme to be learned on the time axis is returned from the output unit 30 to the learning unit 32, and the average value of the LPG cepstral coefficients of the phoneme to be learned is calculated. Repeat the above as many times as necessary while changing the type of voice.

In addition to the original average value (mi j' ) without learning, which is obtained by appropriately weighting the average value for each phoneme, create a new average value for each phoneme and use the average value mi in the coefficient memory 25.
Replace 1'. Furthermore, using this average value, the discriminant coefficient aij and the constant term (second term) of equation (4) are corrected for each phoneme, and these are transferred to the coefficient memory 26 as a new standard pattern, and the standard pattern is rewritten. .

Next, a case in which speech recognition is actually performed will be explained. Regarding the unknown audio signal input from the microphone 10, the signal processing circuit 22 and the linear predictive analysis processor 23 are used to calculate the LPG cepnutrum coefficient "("+y"2+...
・+

This is done for each phoneme (1 = I H2+..., n) (
n is the number of phonemes) and transferred to the main memory 27.

The main processor 28 uses this similarity and the band filter 2
By combining the results of segmentation based on the output of step 6, phoneme recognition is performed and a phoneme sequence is created.

Finally, the phoneme sequence is compared with the word dictionary memory 29, and the word with the highest degree of similarity is transferred to the output unit 30 as a recognition result.

In the above embodiment, before performing voice recognition, a voice whose content is known in advance is input, and based on the result, the coefficient memory 26
I mentioned the case of modifying the standard pattern in
Of course, the standard pattern in the coefficient memory 26 may be modified based on the recognition result of unknown speech during speech recognition.

In this case, there is no need to learn speech whose contents are known in advance, and it is possible to automatically follow changes in the environment, changes in the input person's voice, etc.

As described above, this embodiment is a speech recognition device based on phoneme recognition, and is characterized by having a learning function that creates a standard pattern for each phoneme to suit the user through simple learning in advance, and is capable of achieving high speech recognition. performance can be achieved. Also, the calculations for learning are extremely simple,
without requiring a calculation circuit with special high calculation accuracy.
New standard patterns can be created immediately.

Figure 3 shows a comparison of phoneme recognition rates between 10 adult males with and without learning.

The results were 34 when learning was performed using all words for evaluation and 36 when learning was performed using a small number of words of about 20 words. In both cases, the phoneme recognition rate improved compared to the case without learning33, indicating that it was particularly effective for speakers (NS, KS, SM, etc.) who had traditionally had extremely low recognition rates. .

Figure 4 shows the standard deviation of the recognition rate for each phoneme, and shows the standard deviation of the recognition rate for each phoneme, which is 41 when learning is performed on all words compared to 41 without learning.
2. When performing with minority words, the variation in both 43 is reduced,
This shows that it has a favorable effect on word matching in the later stage.

This embodiment has the following effects.

■ By equipping a speech recognition device with a learning function, it is possible to automatically create a standard pattern suitable for the user and achieve good speech recognition accuracy with less variation due to changes in the environment or individual differences between speakers. can.

■Learning can be performed automatically by uttering a small number of sounds before or during use, and standard patterns can be created extremely easily and quickly without the need for special equipment.

Effects of the Invention In short, the present invention includes an acoustic analysis section that calculates a spectrum or information similar to it (hereinafter referred to as spectral information) from an audio signal divided into each phoneme, and an acoustic analysis unit that calculates a spectrum or information similar to it (hereinafter referred to as spectral information) from an audio signal divided into each phoneme, and an acoustic analysis unit that calculates a spectrum or information similar to it (hereinafter referred to as spectral information). a coefficient storage unit that stores a standard pattern in advance; a discrimination filter unit that uses the spectrum information and the standard pattern to obtain a filter output for each phoneme; and a word dictionary that stores a word dictionary expressed in terms of similarity or phoneme sequence. a storage unit, an output unit that compares the degree of similarity or phoneme sequence created through the discrimination filter unit with a word dictionary and outputs a word with the highest degree of similarity as a recognition result, and a result of the output unit and the acoustic analysis unit. and a learning section that creates a new standard pattern from spectral information and rewrites the contents of the coefficient storage section based on the result, the speech recognition device improves speech recognition accuracy by a speaker. This method has the advantage that it can be used stably for unspecified speakers.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device based on phoneme recognition, Fig. 2 is a block diagram of a speech recognition device according to an embodiment of the present invention, and Fig. 3 shows the effects of the speech recognition device of the present invention. FIG. 4 is a diagram showing the effect of the speech recognition device of the present invention as a standard deviation for each phoneme. 21...AD converter, 22...signal processing circuit, 23...linear predictive analysis processor, 2
4. Old... 4'lJ separate filter, 25... Coefficient memory, 27... Main memory, 28...
・Main processor, 29...Word dictionary memory,
30... Output section, 32. Old... Learning section. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 4
Letter of amendment to figure procedure Dear Commissioner of the Japan Patent Office ■ Display of the case 1982 Patent Application No. 163537 3 Person making the amendment Relationship with the case Patent application Person Address 1006 Kadoma, Kadoma City, Osaka Name Name (
582) Matsushita Electric Industrial Co., Ltd. Representative Toshihiko Yamashita 4 Agent 571 Address 1006 Oaza Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. Specification 1, Name of the invention Voice recognition device 2, Patent claim Scope (1) An acoustic analysis unit that calculates a spectrum or information similar to it (hereinafter referred to as spectral information) from an audio signal, and a coefficient storage unit that stores in advance a standard pattern obtained from a standard audio signal composed of multiple speakers. , a similarity calculation unit that calculates the similarity of each phoneme using the spectral information and the standard pattern; a word dictionary storage unit that stores a word dictionary expressed in similarities or phoneme sequences; and the similarity calculation unit. an output unit that compares the similarity or phoneme sequence created through the process with a word dictionary and outputs the word with the highest degree of similarity as a recognition result; and a new standard from the results of the output unit and the spectrum information of the acoustic analysis unit. A speech recognition device comprising: a learning section that creates a pattern and rewrites the contents of the coefficient storage section based on the result. ('4) The speech recognition device according to claim 1, characterized in that the standard pattern includes at least a variance-covariance matrix and an average value of spectral information. The speech recognition device according to claim 1, characterized in that the content of the speech recognition device is modified. 3. Detailed Description of the Invention Industrial Field of Application The present invention is directed to a speech signal uttered by a human voice This article relates to a speech recognition device that automatically recognizes speech.The structure of conventional examples and their problemsSpeech recognition devices that automatically recognize speech are a means of giving data and commands from humans to electronic computers and various machines. It is considered to be very effective as a method of speech recognition.The operating principle of many speech recognition devices that have been researched or published in the past is the slam matching method.This method can be used for all types of words that need to be recognized. The standard pattern is memorized in advance, and the degree of matching (hereinafter referred to as similarity) is calculated by comparing it with the unknown input button, and it is determined that the word is the same as the standard button that provides the maximum match. In this method of matching, a standard button must be prepared for every word to be recognized, so if the speaker changes, it is necessary to manually memorize a new standard button. Therefore, when there are hundreds of types of words to be recognized, it takes time and effort to pronounce and register all types of words, and the memory capacity required for registration is enormous. Furthermore, the time required to match the input sound with the standard pattern increases as the number of words increases.On the other hand, it is difficult to divide the input speech into phoneme units and recognize them as combinations of phonemes. This method (hereinafter referred to as phoneme recognition) that measures the similarity with a word dictionary written in phoneme units requires significantly less memory capacity for the word dictionary, shortens the time required for bang matching, and allows the contents of the dictionary to be changed easily. It has the feature of being easy.For example, the utterance of "red" is /
a/. AKAI is created by combining the three phonemes /ni/+/li.
Since it can be expressed in an extremely simple form,
It is easy to handle speech with multiple words by any speaker. FIG. 1 shows a block diagram of a speech recognition system characterized by performing phoneme recognition. Audio input through a microphone or the like is analyzed by an acoustic analysis section 1. As an analysis method, a group of bandpass filters and linear predictive analysis are used to obtain spectrum information at every frame period (about 10 mS). The phoneme discrimination unit 2 uses the spectrum information obtained by the acoustic analysis unit 1 to discriminate phonemes for each frame based on the data in the standard pattern storage unit 3. Standard pattern storage section 3
The standard patterns stored in are prepared in advance for each phoneme from the voices of many speakers. The segmentation unit 4 detects speech sections and determines boundaries for each phoneme (hereinafter referred to as segmentation) based on the analysis output of the acoustic analysis unit 1. The phoneme recognition section 6 performs the work of determining what phoneme is for each phoneme section based on the results of the segmentation section 4 and the phoneme discrimination section 2. As a result, a series of phonemes is completed. The word recognition unit 6 compares this phoneme sequence with a word dictionary 7 written in phoneme sequences in a regular manner, and outputs the word with the highest degree of similarity as a recognition result. The most important point when using the above method to target unspecified speakers is to stably obtain high speech recognition accuracy for any speaker environment. Furthermore, this should not place too much burden on the speaker or require expensive components in the speech recognition device. However, speech recognition devices that have been announced or prototyped so far have had the drawback of not meeting the above conditions. As a conventional example, a method that targets prediction residuals (see Kano, ``Evaluation of LPG distance scale for vowel recognition in conversational speech'', Journal of the Institute of Electronics and Communication Engineers 80/5. VOL T-63D, 爲6) ), the maximum parameter of phoneme i is determined in advance by linear predictive analysis from the voices of multiple speakers Am r ()-1t 2 t...
P) (P is the analysis order), and the prediction residual is calculated using the following equation, where Sj is an autocorrelation coefficient calculated from unknown input speech. This prediction residual Ni is used as a distance measure for each target phoneme, and the phoneme with the minimum Ni is determined as the discrimination result. However, in this method, the maximum parameter Ai5, which corresponds to the standard pattern of phonemes, is just an average value, so even if a learning function is provided to recreate A to suit the user, it will not be possible to deal with variations in pronunciation due to articulatory combination. However, the problem was that the recognition rate was low. OBJECTS OF THE INVENTION The present invention solves the above-mentioned drawbacks, and provides a speech recognition device that can deal with unspecified speakers and stably obtain high speech recognition accuracy without being affected by differences in speakers, environments, and language. The purpose is to Structure of the Invention The present invention has been made to achieve the above object, and includes an acoustic analysis section that calculates a spectrum or information similar to it (hereinafter referred to as spectrum information) from an audio signal, and a standard audio signal composed of multiple speakers. a coefficient storage unit that stores in advance the standard pattern obtained from the above, a similarity calculation unit that calculates the similarity of each phoneme using the spectral information and the standard pattern, and a word dictionary in which the similarity or phoneme sequence is expressed. an output unit that compares the similarity or phoneme sequence created through the similarity calculation unit with a word dictionary and outputs the word with the highest degree of similarity as a recognition result; The apparatus includes a learning section that creates a new standard pattern from the result and spectrum information from the acoustic analysis section and rewrites the contents of the coefficient storage section based on the result. DESCRIPTION OF EMBODIMENTS FIG. 2 shows an embodiment of the configuration of a speech recognition apparatus according to the present invention. The audio signal received from the microphone 31 is sent to the AD converter 21.
12 songs are sampled and converted to 12 bit. A signal processing circuit applies pre-emphasis and a Hamming window of 2Qm3 to this signal, and a linear predictive analysis processor 23 processes it every 10 mS.
Calculate the LPC cepstral coefficients. The LPC cepstral coefficients are passed through the similarity calculation unit 24 to calculate the similarity for each phoneme for each frame and transferred to the main memory 27. Coefficient memory 26 stores filter coefficients for each phoneme. On the other hand, the bandpass filter 26 calculates band powers and total powers of about three channels, and transfers them to the main memory 27 as data for phoneme segmentation. The main processor 28 uses the results of the similarity calculation unit 24 and the bandpass filter 26 to detect speech intervals and perform segmentation for each phoneme, and then selects the phoneme with the highest degree of similarity from the similarity of each phoneme in the similarity calculation unit 24. is determined for each section and a phoneme sequence is created. By comparing this phoneme sequence with the word dictionary memory 29 which is also written in a phoneme sequence, the word name with the highest degree of similarity is outputted to the output unit 3o as a recognition result. However, although this alone can be used for unspecified speakers, since the coefficient memory 26 that corresponds to the standard pattern is fixed, the recognition performance varies greatly depending on the speaker.
There are cases where the recognition rate becomes considerably low. Therefore, a learning section 32 is provided to provide a new learning function. This learning section 32 receives the LPC cepstral coefficients obtained by the linear predictive analysis processor 23, creates learning data with reference to the results obtained from the output section 3o, and creates learning data based on the variance and covariance matrix prepared in advance. Recalculate the discriminant coefficient for each phoneme that is most suitable for the speaker and memo the coefficient! 725. Next, the operation of the phoneme recognition device according to the present invention will be explained in detail with reference to FIG. The vowel /a is converted from the many word sounds uttered by many speakers inputted in advance from the microphone 31 through the AD converter 21.
The nasal sounds /, 10/, /u/, /V, /e/ are cut out in advance. Using this audio data, the signal processing circuit 22 and the linear predictive analysis processor 23 perform linear predictive analysis for each analysis section of 10 m5 to calculate p-dimensional LPC Keb Nutram coefficients. A covariance matrix W for all phonemes and an average value tnt (i represents the type of phoneme) for each phoneme are calculated using Jin's LPG cepstral coefficients. From this result, the discrimination coefficient aij (j=1.2
．． . . . tp) can be expressed as follows, assuming that the (i, j') element of the inverse matrix W-1 of the covariance matrix W is June. For each phoneme, ai 5, m, ,', δth, m1tW-
'm, (Nukimichi) as standard pattern coefficient memory 25
Store it in Next, have the user utter a voice whose content is somewhat distorted (e.g. /a/, /V, /u/, /@/, 10/), and create an LPC cepstrum for each analysis interval in the voice interval. The coefficients are calculated by the linear prediction analysis processor 23 and transferred to the learning section 32. On the other hand, the degree of similarity is determined by the discrimination filter 24 using the standard pattern stored in the coefficient memory 26 that is stored in advance. The similarity calculation unit 24 can express the Mahalanobis distance with respect to the LPC cepstral coefficient Li is simply expressed as 0, and the similarity is calculated using equation 4.0, and the result is stored in the main memory 2.
7 and creates a phoneme sequence through the main processor 28. Next, a value indicating the position of the phoneme to be learned on the time axis is returned from the output unit 3o to the learning unit 32, and the average value of the LPC cepstral coefficients of the phoneme to be learned is calculated. Repeat the above as many times as necessary while changing the type of voice. In addition to the original average value (m,') in the case of not learning, which is obtained by appropriately weighting the average value for each phoneme, a new average value for each phoneme is created and the average value m157 is replaced in the coefficient memory 26. Furthermore, using this average value, the discrimination coefficient ai
The constant term (second term) of equations l and (4) is corrected for each phoneme, and Nire et al. is transferred to the coefficient memory 25 as a new standard pattern, and the standard pattern is rewritten.Next, speech recognition is actually performed. A case will be described. For an unknown audio signal input from the microphone 10, the signal processing circuit 22 and the linear predictive analysis processor 23 are used to calculate the LPC cepstral coefficients x(xl, x2...
・t every (l = 1, 2..., n) (n is the number of phonemes) and transfers it to the main memory 27.The main processor 28 uses this similarity and the band filter 2.
By combining the results of segmentation based on the output of step 6, fortune-telling phoneme recognition is performed and a phoneme sequence is created. Finally, the phoneme sequence is compared with the word dictionary memory 29, and the word with the highest degree of similarity is transferred to the output unit 30 as a recognition result. In the above embodiment, before performing voice recognition, a voice whose content is known in advance is input, and based on the result, the coefficient memory 25
I mentioned the case of modifying the standard pattern in
Of course, the standard pattern in the coefficient memory 25 may be modified based on the speech recognition result during speech recognition. In this case, there is no need to previously learn speech with strange content, and it is possible to automatically follow changes in the environment, changes in the input person's speech, etc. As described above, this embodiment is a speech recognition device based on phoneme recognition, and is characterized by having a learning function that creates a standard pattern for each phoneme to suit the user through simple learning in advance, and is capable of achieving high speech recognition. performance can be achieved. Also, the calculations for learning are extremely simple,
without requiring a calculation circuit with special high calculation accuracy.
You can quickly create new standard batons. Figure 3 shows a comparison of phoneme recognition rates between 10 adult males with and without learning. The results were 34 when learning was performed using all words for evaluation and 36 when learning was performed using a small number of words of about 20 words. In both cases, the phoneme recognition rate improved compared to the case without learning33, indicating that it is particularly effective for speakers who previously had extremely low recognition rates (NS, KS, SM, etc.). . Figure 4 shows the standard deviation of the recognition rate for each phoneme, and shows the standard deviation of the recognition rate for each phoneme, which is 41 when learning is performed on all words compared to 41 without learning.
2. When performing with minority words, the variation in both 43 is reduced,
This shows that it has the effect of improving the performance of word matching in the subsequent stage. This embodiment has the following effects. - By equipping the speech recognition device with a learning function, it is possible to automatically create standard batons suited to the user and achieve good speech recognition accuracy with little variation due to changes in the environment or individual differences among speakers. ■Learning can be performed automatically by uttering a small number of sounds before or during use, and standard patterns can be created extremely easily and quickly without the need for special equipment. Effects of the Invention In short, the present invention includes a coefficient storage section that stores in advance a standard pattern obtained from a standard speech signal that is a spectrum of a speech signal, and a coefficient storage section that stores in advance a standard pattern obtained from a standard speech signal that is a spectrum of a speech signal, and a coefficient storage section that measures the similarity of each phoneme using the spectral information and the standard pattern. a word dictionary storage unit that stores a word dictionary expressed in similarities or phoneme sequences; and a word dictionary storage unit that stores a word dictionary expressed in similarities or phoneme sequences, and compares the similarity or phoneme sequence created through the similarity calculation unit with the word dictionary to find the most an output unit that outputs high-value words as recognition results; and a learning unit that creates a new standard pattern from the results of the output unit and the spectrum information of the acoustic analysis unit and rewrites the contents of the coefficient storage unit based on the results. The present invention provides a speech recognition device having the following features, which has the advantage of greatly reducing variations in speech recognition accuracy between speakers and being able to be used stably for unspecified speakers. 4. Brief description of the drawings Fig. 1 is a block diagram of a conventional speech recognition device based on phoneme recognition, Fig. 2 is a block diagram of a speech recognition device according to an embodiment of the present invention, and Fig. 3 is a block diagram of a speech recognition device according to an embodiment of the present invention. FIG. 4 is a diagram showing the effectiveness of the speech recognition device of the present invention for each speaker, and FIG. 4 is a diagram showing the effectiveness of the speech recognition device of the present invention as a standard deviation for each phoneme. 21...AD converter, 22...signal processing circuit, 23...linear predictive analysis processor, 2
4...Similarity calculating unit, 26...Coefficient memory, 27...Main memory, 28...Main processor, 29...Word dictionary memory ,3
0 songs...How to play part, 32...Learning part. Name of agent: Patent attorney Toshio Nakao, 1st person, 2nd person
Diagram No.

Claims (1)

[Scope of Claims] (1) An acoustic analysis unit that calculates a spectrum or information similar to it (hereinafter referred to as spectrum information) from a speech signal divided into phonemes, and a coefficient storage section that stores in advance a standard pattern based on the spectral information; a discrimination filter section that uses the spectral information and the standard pattern to obtain a filter output for each phoneme; and a word dictionary that stores a word dictionary expressed in terms of similarity or phoneme sequence. a dictionary storage unit, an output unit that compares the degree of similarity or phoneme sequence created through the discrimination filter unit with a word dictionary and outputs a word with the highest degree of similarity as a recognition result, and a result of the output unit and the acoustic analysis. 1. A speech recognition device comprising: a learning section that creates a new standard pattern from spectrum information of the coefficient storage section and rewrites the contents of the coefficient storage section based on the result. (Rumor) The speech recognition device according to claim 1, characterized in that the standard pattern includes at least a variance-covariance matrix and an average value of spectral information. (3) Coefficient storage based on the recognition result of unknown input speech. 2. The speech recognition device according to claim 1, wherein the content of the part is modified.