JPH042197B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device for automatically recognizing an audio signal 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 instructions from humans to computers and various machines.

The pattern 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 unknown input patterns. The word is calculated and determined to be the same word as the standard pattern that yields the maximum match. In this pattern matching method, standard patterns must be prepared for all words to be recognized, so if the utterance changes, a new standard pattern must be input and stored. Therefore, if hundreds of 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. be done. Furthermore, there is a drawback that the time required for pattern matching between the input pattern and the standard pattern increases as the number of words increases.

On the other hand, input speech is divided into phoneme units and recognized as combinations of phonemes (hereinafter referred to as phoneme recognition).
The method of determining similarity with a word dictionary written in phoneme units has the advantage that the memory capacity required for the word dictionary is significantly smaller, the time required for pattern matching is shorter, and the contents of the dictionary can be easily changed. ing. For example, the utterance of "red" is /
Since the three phonemes a/, /K/, and /i/ can be combined and expressed in an extremely simple format called AKAI, it is easy for non-specific speakers to deal with the sounds of many words.

FIG. 1 shows a block diagram of a speech recognition system characterized by phoneme recognition. Audio input through a microphone or the like is analyzed by the acoustic analysis section 1.
The analysis method uses a group of bandpass filters and linear predictive analysis, and spectral information is obtained every frame period (about 10 mS). The phoneme discrimination section 2 uses the spectrum information obtained by the acoustic analysis section 1 and performs phoneme discrimination for each frame based on the data in the standard pattern storage section 3. The standard patterns stored in the standard pattern storage section 3 are obtained in advance for each phoneme from the voices of multiple speakers. The segmentation unit 4 detects speech intervals 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 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 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. Also,
This must not place too much burden on the speaker or require expensive parts for 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 (Kano, 2003, ``For the purpose of vowel recognition in conversational speech'')
“Evaluation of LPC distance scale” Journal of the Institute of Electronics and Communication Engineers 80/5,
VOLJ-63D, No. 5), the maximum parameter A _ij (j = 1, 2, ..., P) (P is the analysis order) of phoneme i is determined in advance by linear predictive analysis from the voices of many speakers. Then, calculate the prediction residual using the following formula.

N _i = _p 〓 ^j=1 A _ij S _j …(1) Here, S _j is an autocorrelation coefficient obtained from unknown input speech. This prediction residual N _i is obtained for each target phoneme and is used as a distance measure, and the phoneme with the minimum N _i is taken as the discrimination result.

However, in this method, the maximum parameter A _ij corresponding to the standard pattern of phonemes is just an average value, so
Even if a learning function was provided to recreate A _ij to suit the user, it would not be able to deal with variations in utterances caused by articulatory combinations, resulting in a low recognition rate.

Purpose of the Invention The present invention solves the above-mentioned drawbacks, and provides a speech recognition device that can handle 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 similarity calculation unit that calculates similarity for each phoneme using a standard pattern that includes at least the variance/covariance matrix and average value of spectral information obtained from a word dictionary storage unit for storing a word dictionary; an output unit that compares the degree of similarity or phoneme sequence created through the similarity calculation 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 a learning section that creates a new average value from the spectral 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 device according to the present invention. The audio signal input from microphone 31 is AD
A converter 21 converts it to 12 bits with 12kHz sampling. A signal processing circuit applies pre-emphasis and a 20 mS Hamming window to this signal, and a linear predictive analysis processor 23 calculates LPC cepstrum coefficients every 10 mS. 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. The coefficient memory 25 stores filter coefficients for each phoneme.

On the other hand, the band filter 26 calculates the band power and total power of about 3 channels, and stores it in the main memory as data for phoneme segmentation.
Transfer to 7. The main processor 28 uses the results of the similarity calculation unit 24 and the band filter 26 to detect speech intervals and perform segmentation for each phoneme, and then calculates the similarity from the similarity for each phoneme in the similarity calculation unit 24. The highest phoneme 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 30 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, recognition performance varies greatly depending on the speaker, and the recognition rate becomes considerably low. A situation arises. Therefore, a learning section 32 is provided to provide a new learning function. This learning section 32 uses the LPC obtained by the linear predictive analysis processor 23.
Receiving the cepstral coefficients, creating learning data with reference to the results obtained from the output unit 30, and recalculating the discriminant coefficient for each phoneme that is most appropriate for the speaker based on the variance and covariance matrix determined in advance. , performs an operation for transferring 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.

Vowels /a/, /o/, /u/, /i/, / are passed through the AD converter 21 from a large number of word sounds uttered by multiple speakers inputted in advance from the mask 31.
Cut out e/ and nasal sounds. 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 LPC cepstral coefficients. Using these LPC cepstral coefficients, a covariance matrix W for all phonemes and an average value m _i (i represents the type of phoneme) for each phoneme are determined. From this result, the discriminant coefficient a _ij for phoneme i
(j = 1, 2, ..., p) is the inverse matrix W ^-1 of the covariance matrix W. Let δ ^jj ' be the (j, j') element, then a _ij = 2 _p 〓 ^j=1 δ ^jj ' It can be expressed as m _ij ′ …(2).

For each phoneme, a _ij , m _ij ', δ ^ij ', m _i ^t W ^-1 m _i (post-reverse) are obtained and stored in the coefficient memory 25 as a standard pattern.

Next, the user can hear a voice whose content is known in advance (e.g. /a/, /i/, /u/, /
e/, /o/) is uttered, the linear predictive analysis processor 23 calculates the LPC cepstral coefficients for each analysis section in the speech section, and transfers them 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 25 that is stored in advance. In the similarity calculation unit 24, the Mahalanobis distance D _i ² for the LPC cepstral coefficient x of the input signal is D _i ² =x ^t W ^-1 x− _p 〓 ^j=1 a _ij x _j +m _i ^t W ^-1 m _i …( 3) (t indicates the transposed matrix) However, since the first term does not depend on the type of phoneme, the similarity L _i can be simply expressed as L _i = _p 〓 ^j=1 a _ij x _j − m _i ^t W ^-1 m _i ...(4), and the similarity is calculated using equation (4). The results are transferred to the main memory 27 and passed through the main processor 28 to create a phoneme sequence. Next, the 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 value indicating the position of the phoneme to be learned on the time axis is returned to the learning unit 32.
Find the average value of the LPC cepstral coefficients. Repeat the above as many times as necessary while changing the type of voice.
In addition to the original average value (m _ij ′) without learning, the average value for each phoneme is appropriately weighted,
A new average value for each phoneme is created and the average value m _ij ' in the coefficient memory 25 is replaced. Furthermore, this average value is used to calculate the discriminant coefficient a _ij and the constant term (second term) in equation (4).
is modified for each phoneme, and these are transferred to the coefficient memory 25 as a new standard pattern, and the standard pattern is rewritten.

Next, a case in which speech recognition is actually performed will be explained. For 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 LPC cepstral coefficients x (x ₁ , x ₂ , ..., x _p ), and the similarity calculation unit 2
4, and using a standard pattern previously determined and stored in the coefficient memory 25, the degree of similarity L _i of phoneme i is determined from equation (4).

This is obtained for each phoneme (i=1, 2, . . . , n) (n is the number of phonemes) and transferred to the main memory 27.
The main processor 28 performs phoneme recognition by combining this degree of similarity and the result of segmentation based on the output of the band filter 26 to create a phoneme sequence.

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 voice recognition, a voice whose content is known in advance is input, and the standard pattern in the coefficient memory 25 is corrected based on the result. Of course, the standard pattern in the coefficient memory 25 may be modified based on the recognition result.

In this case, there is no need to learn speech whose content is 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 speech recognition, and is characterized by having a learning function that creates standard patterns for each phoneme to suit the user through simple learning in advance. It can have voice recognition performance. In addition, calculations for learning are extremely simple, and new standard patterns can be created immediately without requiring a special calculation circuit with high calculation accuracy.

Figure 3 shows a comparison of phoneme recognition rates between 10 adult males with and without learning. When learning is performed using all words for evaluation
34, and 35 when using a minority of around 20 words. In both cases, the speech recognition rate improved compared to the case without learning33, and it was shown that it is particularly effective for speakers (NS, KS, SM, etc.) for whom recognition rates have traditionally been extremely low. There is.

Figure 4 shows the standard deviation of the recognition rate for each phoneme. Compared to the case without learning41, the variation is reduced both when learning is performed on all words42 and when it is performed on a minority of words43. 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 a speech recognition device with a learning function, it is possible to automatically create a standard pattern suitable for the user, and to 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 summary, the present invention includes an acoustic analysis unit that calculates a spectrum or information similar to it (hereinafter referred to as spectrum information) from an audio signal, and a dispersion and analysis of the spectrum information obtained from a standard audio signal composed of multiple speakers. a coefficient storage unit that stores in advance a standard pattern including at least a covariance matrix and an average value; a similarity calculation unit that calculates a similarity for each phoneme using the spectral information and the standard pattern; a word dictionary storage unit that stores the word dictionary that has been created; an output unit that compares the similarity or phoneme sequence created through the similarity calculation unit with the word dictionary and outputs the word with the highest degree of similarity as a recognition result; Provided is a speech recognition device, comprising a learning section that creates a new average value from the result of the output section and spectrum information of the acoustic analysis section, and rewrites the contents of the coefficient storage section based on the result. This method has the advantage of significantly reducing variations in speech recognition accuracy between speakers and being able to be used stably for unspecified speakers.

[Brief explanation of 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 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 prediction analysis processor, 24...Similarity calculation unit, 25...Coefficient memory, 27...Main memory, 28...Main processor, 29...Word dictionary memory, 30...Output unit, 32... ...Learning Department.

Claims (1)

[Claims] 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 an acoustic analysis unit that calculates the variance/covariance matrix and average value of the spectral information obtained from a standard audio signal composed of multiple speakers. a coefficient storage section that stores in advance standard patterns including at least one; a similarity calculation section that calculates a degree of similarity for each phoneme using the spectral information and the standard pattern; and a dictionary of words expressed in degrees of similarity or phoneme sequences. a word dictionary storage unit; 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 similarity as a recognition result; A speech recognition device comprising: a learning section that creates a new average value from the spectrum information of the acoustic analysis section and rewrites the contents of the coefficient storage section based on the result.