JPH0344320B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a voice recognition method for automatically recognizing voice 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 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. This method memorizes standard patterns for all types of words that need to be recognized in advance, and compares them with unknown input patterns to calculate the degree of matching (hereinafter referred to as similarity). 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 speaker changes, a new standard pattern must be input and stored. Therefore, this method has a simple principle and is an effective method for recognizing a small number of words, but in cases where more than several hundred types of words are to be recognized, it is difficult to pronounce and register all types of words. It is expected that this will require time and effort, and that the memory capacity required for registration will also be enormous.
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 speech of many words.

FIG. 1 shows a block diagram of a speech recognition method characterized by performing phoneme recognition. Audio input through a microphone or the like is analyzed by the acoustic analysis section 1.
As an analysis method, consider band filter group or linear predictive analysis, and 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. The standard patterns stored in the standard pattern storage section 3 are obtained in advance for each phoneme from the voices of many speakers. In the segmentation section 4,
Based on the analysis output of the acoustic analysis unit 1, voice sections are detected and boundaries for each phoneme are determined (hereinafter referred to as segmentation). The phoneme recognition unit 5 performs the work of determining what phoneme is for each phoneme interval based on the results of the phoneme discrimination unit 2 of the segmentation unit 4. As a result, a series of phonemes is completed. The word recognition unit 6 compares this phoneme sequence with a word dictionary 7 that is similarly expressed 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 or environment.
Furthermore, this should not place too much burden on the speaker or require expensive parts in the case of a 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 targeting prediction residuals (Kano, Koda, "Evaluation of LPC distance measure for vowel recognition in conversational speech", Journal of the Institute of Electronics and Communication Engineers 80/5,
VOLJ-63D, No. 5), the maximum likelihood parameter A _ij (j = 1, 2, ..., p) of phoneme i (j = 1, 2, ..., p) (p is the analysis order) is calculated in advance by linear predictive analysis from the voices of many speakers. is calculated in advance, and the predicted residual is calculated using the following formula.

N _i = _P 〓 ^j=1 A _ij S _j where 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 for which N _i is the minimum is taken as the discrimination result.

However, in this method, the maximum likelihood 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.

In addition, since phonemes such as vowels and semi-vowels are discriminated using standard patterns in frame units, and segmentation and phoneme recognition are performed as a combination of the discrimination results, temporal functions cannot be fully captured and the recognition level does not improve. There was a drawback.

Purpose of the Invention The present invention solves the above-mentioned drawbacks, provides a speech recognition method that can cope with unspecified speakers, and can stably obtain high speech recognition accuracy without being affected by differences in speakers or words. The purpose is to

Structure of the Invention The present invention achieves the above object by creating, for each phoneme, a standard pattern consisting of LPC cepstral coefficients of a plurality of frames expressing temporal movement within a phoneme from the speech of multiple speakers. The present invention provides a speech recognition method characterized in that speech recognition is performed by similarity or phoneme sequence obtained based on a statistical distance measure using a standard pattern and LPC cepstral coefficients of a plurality of frames of unknown speech. be.

DESCRIPTION OF EMBODIMENTS The speech recognition method of the present invention utilizes the fact that there is a difference in the time-frequency characteristics of the spectra of vowels and semi-vowels, or vowels.

An example of how to recognize vowels and semi-vowels will be explained.

Figure 2 A shows the vowel when saying OOSAMA/
Figure 2B shows the spectrum of the transition from the semi-vowel /j/ to the vowel /a/ in the /ja/ part when uttering YASUMONO. The vertical axis represents frequency and the horizontal axis represents time.
Expressed in frames of 10mS. Also, the horizontal axis simultaneously represents the spectral intensity. The peak to the right of the spectrum represents the formant, and the movement of the formant is shown by a broken line.

Comparing A and B, the difference between a vowel and a semi-vowel is that the function of the formant, which moves from the beginning of the phoneme to the center of the vowel (in this case /a/) (shown in Figure A), is different. That is, the time to the center of the phoneme for /a/ is 5 frames, while for /ja/ it is 10 frames. Furthermore, the formant position of /a/ is between 500 and 1000 Hz, whereas the position of /ja/ is widely between 250 and 1000 Hz.

This tendency is also the same between vowels.

For example, even when comparing /a/ and /o/, the state of the spectrum change toward the center and away from the center is different.

Focusing on this phenomenon, the present invention constructs a standard pattern of vowels and semi-vowels by frequency spectra of multiple frames, and compared to the conventional method of distinguishing phonemes in frame units, /a/, /i/, /u/,/
This is intended to accurately distinguish between phonemes such as e/, /o/, /j/, /w/, etc.

In order to obtain high speech recognition accuracy, distance measures in the present invention include distances based on Bayesian judgment,
Statistical distance measures such as Mahalanobis' general distance, linear discriminant function, etc. are preferred.

From the perspective of reducing the amount of calculation, it is desirable to use the Mahalanobis general distance as a basis and expand it as a first-order discriminant function (referred to as the simplified Mahalanobis distance). In the following example, the simplified Mahalanobis distance is used as an example. Let me explain the case. Mahalanobis' generalized distance requires matrix calculations, but if the variances between target phonemes are not significantly different, the co-fraction matrix can be shared, and it can be expanded to a simplified Mahalanobis distance that requires less calculation. I can do it.

Based on the above idea, an embodiment of the speech recognition method according to the present invention will be described with reference to FIG.

First, the speech corresponding to the phoneme i of the majority of speakers for standard pattern creation is input, and in block 11, it is analyzed as spectrum information with an analysis order p and a frame number n.
A two-dimensional array of LPC cepstral coefficients consisting of a plurality of frequency spectra is obtained.

〓i=C _i11 , C _i12 , ..., C _i1p C _i21 , C _i22 , ..., C _i2p 〓 C _io1 , C _io2 , ..., C _iop This is converted into M=n×P-dimensional vector〓 _i in block 12 Make it.

〓 _i = (C _i11 , C _i12 , ..., C _i1p , C _i21 , ..., C _i2
p , _Cio11 ,..., _Ciop ) Using this _〓i , a standard pattern for phoneme i is created in block 13. The above steps are performed for each phoneme.

If you need to learn next time, go to block 14.
The system learns the user's voice and modifies the standard pattern. Learning processes can be established as needed.

On the other hand, the unknown voice is analyzed in block 15 and an LPC consisting of analysis order p and frame number n is used.
Find the two-dimensional array 〓 of the Kesp Slam coefficients.

〓 _i = C ₁₁ , C ₁₂ , ..., C _1p C ₂₁ , C ₂₂ , ..., C _2p 〓 _Co1 , _Co2 , ..., C _op This is converted into an M-dimensional vector 〓 in block 16.

〓=(C ₁₁ , C ₁₂ ,..., C _1p , C ₂₁ ,..., C _2p , _Co1
,..., C _op ) Using this 〓 and standard pattern, block 17
For each phoneme, the degree of similarity is determined using the simplified Mahalanobis distance, and discrimination is performed.

Based on the above idea, FIG. 4 shows the procedure for creating standard patterns for vowels and semi-vowels in the case of learning as an example. Linear predictive analysis is performed on the input speech corresponding to phoneme i in block 21, and a two-dimensional pattern is constructed using n LPC cepstral coefficients C _ij1 to C _ijp to be used as a standard pattern as a frequency axis. i represents the type of phoneme, and j represents the frame order. p is the analysis order. Next block 2
Rearrange the parameters in step 2 and get 〓 _i = (C _i11 , C _i12 ,..., C _i1p , C _i21 ,..., C _i3
1 , ..., C _io1 , C _iop ).

Furthermore, in block 23, a large number of voices 〓 _i
are totaled, and the average value of 〓 _i is set as 〓 _i (〓 _i is an M-dimensional vector). The covariance matrix obtained in block 24 is the same regardless of the type of phoneme, and is represented by . In block 25, set the inverse matrix of 〓 to 〓 ^-1 ,
When the (j, j') elements are set to σj, j' in block 26, the discriminant coefficient a _ij for the j-th parameter of phoneme i is determined in block 27 as a _ij =2 _M 〓 ^j ′ ⁼¹ σ _ij ′m _ij ′ (1) Here, m _ij ′ is the j′-th component of 〓 _i .

On the other hand, the constant d _i determined by the phoneme is
8, d _i =〓 _i ^t 〓 ^-1 〓 _i (2). Here, t represents a transposed matrix.

The a _ij and d _i obtained in the above manner are stored in the coefficient memory shown in block 29 as a phoneme standard pattern.

Furthermore, the average value 〓 _i obtained in block 23 and the inverse matrix 〓 ^-1 obtained in block 25 are stored in the learning section shown in block 30 for use in learning.

The standard pattern created as described above is modified through learning. Thereafter, the degree of similarity between the unknown input speech and the modified standard pattern is calculated using the simplified Mahalanobis distance, and phoneme discrimination is performed.

Input parameters = (x ₁ , x ₂ , ..., x _o )
The Mahalanobis distance D _i ² for the distribution of phoneme i is expressed as D _i ² =〓 ^t 〓 ^-1 〓− _M 〓 ^j=1 a _ij χ _j +〓 _i ^t 〓 ^-1 〓 _i (3).

t represents a transposition matrix.

Since the first term in equation (3) does not depend on the type of phoneme,
The degree of similarity L _i can be simply expressed as L _i = _M 〓 ^j=1 a _ij χ _j −〓 _i ^t −〓 ^-1 〓 _i (4). The second term in equation (4) is a constant determined by the type of phoneme, and if this is expressed as d _i based on equation (2), the similarity L _i is L _i = _M 〓 ^j=1 a _ij It can be obtained by χ _j −d _i (5). Here, a _ij and d _i are those already determined as standard patterns.

Equation (5) is called the simplified Mahalanobis distance.

The case of learning and the recognition procedure will be explained using the block diagram of the apparatus shown in FIG. First, the standard patterns a _ij and d _i created by the procedure described in FIG. 4 are stored in the coefficient memory 31. Also, 〓, 〓 ^-1 are stored in the learning section 32.

Next, we will discuss the modification of standard patterns through learning.

That is, the user generates vowels and semi-vowels while facing the microphone 33, A/D converter 34 performs A/D conversion, a signal processing circuit 35 applies a Hamming window, and pre-emphasis is performed. The LPC cepstrum coefficients are determined by the linear predictive analysis processor 36, the coefficients are rearranged, and the coefficients are transferred to the learning unit 32 as χ _i . In this case, parameter information from the band filter 39 may be used as necessary. Furthermore, χ _i is also transferred to the similarity calculation unit 37. The similarity calculating section 37 calculates the similarity L _i shown in the above equation (5) using the standard pattern in the coefficient memory 31.

This is transferred to the main memory 38. On the other hand, parameters for segmentation (band power and total power) are determined by the band filter 39 and transferred to the main memory 38. The main processor 40 determines the position on the time axis to be learned from the result of the similarity calculation section 37 and the result of the band filter 39, and specifies it to the learning section 32 through the output section 41. The learning unit 32 uses the stored average value 〓 _i ,
Using the inverse covariance matrix 〓 ^-1 , perform speaker matching of the standard pattern in the following steps. If the average value to be sought is 〓′ _i , then 〓 _i ′=(α〓 _i +〓 _i )/(α+1) (6). Here α is a weighting coefficient.

Using this 〓′ _i , a′ _ij to be adapted is obtained from equation (1) as follows: a′ _ij =2 _M 〓 ^j=1 σ _ij ′m′ _ij ′ (7). However, m′ _ij ′ is the j′-th component of 〓′ _i .

In addition, the d _i ′ to be matched becomes d′ _i =〓′ _i ^t 〓 ^-1 〓′ _i (8) using equation (2), and these a′ _ij and d′ _i are used as coefficients as a standard pattern adapted to the speaker. Rewrite the memory 31.

This completes the learning, and the actual speech recognition is performed using the following steps.

The input voice is passed through an AD converter 34 and a signal processing circuit 35, and then a linear predictive analysis processor 36 performs a linear predictive analysis to obtain P LPC cepstral coefficients.
Find an M-dimensional input vector 〓 by rearranging the LPC cepstral coefficients for n frames. This χ
Using the modified standard patterns a' _ij and d' _i stored in the coefficient memory 31, the discrimination filter 37 calculates the degree of similarity L _i using the following equation.

L _i = _M 〓 ^j=1 a′ _ij χ _j −d′ _i (9) where χ _j is the j-th component of the input vector 〓.

This L _i and the result of the band filter 39 are transferred to the main memory 38 . The main processor 40 uses these data to perform voice section detection, segmentation, and phoneme recognition to create a phoneme sequence. This phoneme sequence is compared with a word dictionary memory 42 which is also written in a phoneme sequence, and the word name with the highest degree of similarity is outputted to the output unit 41 as a recognition result.

As described above, the method according to the present embodiment is a speech recognition method based on phoneme recognition, in which a standard pattern of a phoneme is composed of frequency spectral frames of multiple frames, thereby detecting temporal movement within a phoneme. Through careful consideration and further learning, phoneme standard patterns can be automatically created and matched to the speaker, resulting in high speech recognition performance. In addition, since Mahalanobis' general distance is used as a distance measure and further simplified, calculations for phoneme similarity and learning are easy, and can be achieved without requiring a calculation circuit with a high calculation precision. I can do it.

FIG. 6 shows the recognition rate of vowels and nasals for 10 adult females using the conventional method of phoneme discrimination in frame units without learning (51), with learning (52), and the method 53 of this embodiment. This is a comparison. If the discrimination is performed on a frame-by-frame basis, it will be considerably improved, but it will not be as effective as this embodiment.

By configuring the standard pattern with frequency spectra of multiple frames in this way, significant improvements are made for all speakers, and the average error rate is
It is 6.7%, which is compressed to 70% of 52 with learning based on frame-by-frame discrimination.

FIG. 7 compares the recognition rate of semi-vowels between the conventional case 61 with frame-based discrimination and learning and the method 62 of this embodiment. When recognized with the conventional method, only 68.5% was achieved on average.
Using this example, it improved to 84.4%. The recognition rate is
It is improved by 15.9% and the error rate can be reduced to 1/2.

In summary, the features of this embodiment are as follows.

(1) A high phoneme recognition rate can be obtained by constructing a standard pattern of phonemes using spectra of multiple frames or information similar to them.

(2) By automatically creating standard phoneme patterns through learning and matching them to the speaker, it is possible to recognize speakers who were previously unrecognizable (for example, speakers YM and YI in Figure 6).
Accurate phoneme recognition is possible even for

(3) Due to the effects of (1) and (2), a high-performance speech recognition device can be constructed, and a high word recognition rate can be expected.

Although the above embodiment describes the case where learning is involved, the feature of the present invention is that a standard pattern consisting of LPC cepstral coefficients of a plurality of frames expressing temporal movement within a phoneme is constructed, and a standard pattern is constructed based on a statistical distance measure. Since the degree of similarity is calculated based on the method, a good phoneme recognition rate can be obtained even when no learning is performed.

Effects of the Invention In summary, the present invention creates a standard pattern for each phoneme consisting of LPC cepstral coefficients of a plurality of frames to express temporal movement within a phoneme from the voices of multiple speakers, and The present invention provides a speech recognition method that performs speech recognition based on the similarity or phoneme sequence obtained based on a statistical distance measure using LPC cepstral coefficients of multiple frames of unknown speech. It is possible to obtain a phoneme recognition rate and realize a speech recognition device with excellent performance.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a conventional speech recognition method based on phoneme recognition, Fig. 2 is a diagram showing an example of a spectral pattern related to the method of the present invention, and Fig. 3 is a diagram according to an embodiment of the present invention. FIG. 4 is a block diagram showing the method of creating a standard pattern of the present invention. FIG. 5 is a block diagram showing an example of the configuration of a speech recognition device embodying the speech recognition method of the present invention. 6 and 7 are diagrams showing the effects of this embodiment in terms of the speech recognition rate for each speaker. 31...Coefficient memory, 32...Learning section, 36...
. . . Linear prediction analysis processor, 37 . . . Similarity calculation section, 38 . . . Main memory, 39 . . . Bandwidth filter, 40 .

Claims (1)

[Claims] 1 Create a standard pattern for each phoneme consisting of LPC cepstral coefficients of multiple frames expressing temporal movement within a phoneme from the voices of multiple speakers,
A speech recognition method characterized in that speech recognition is performed based on a similarity or a phoneme sequence obtained based on a statistical distance measure using the standard pattern and LPC cepstral coefficients of a plurality of frames of unknown speech.