JPH0344317B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD The present invention relates to a speech recognition method for automatically recognizing the content of speech. Configuration of conventional examples and their problems In recent years, research and development on speech recognition for unspecified speakers and multiple words has become active. Conventional phoneme discrimination in speech recognition, which is characterized by phoneme recognition, is described in Ide et al., "Study of consonant recognition using dynamic features of spectrum," Proceedings of the Acoustical Society of Japan, 1981, 10, 2-1-2. It has been stated. The flowchart is shown in Figure 1. First, the procedure for creating a standard pattern will be described.
Every 10mS for consonants and semi-vowels in speech
Obtain the output of a group of 29 channel band filters (Q = 6, center frequency 250-6300 Hz, 1/6 octave interval). Furthermore, several consecutive bands along the frequency axis are collectively defined as 6 channels. These six channels are taken as one frame, and five consecutive frames are extracted to form a 6×5=30-dimensional vector. This is totaled for each phoneme, and the average value m _i for phoneme i,
Let the covariance matrix be W _i and the inverse matrix W ^-1 _i , then m _i
and W ^-1 _i are stored in the standard pattern storage section in advance as standard patterns. Next, the input unknown voice is acoustically analyzed, and the output of the 29-channel band filter X″ ₁ (X″ _1,1 ,X″ _1,2 ,...,X″ _{1 . _ _
. _ _}
Calculate X′ ₃ , _X ′ ₄ ,
= 30-dimensional vector X (X ₁ , X ₂ , ..., X ₃₀ ). Further, for this vector, similarity calculation is performed by Bayesian judgment using the aforementioned standard pattern as shown in Process E. The degree of similarity P _i for phoneme i can be calculated using the following equation. P _i = (2π) ^-15/2 |W _i | ^-1/2 exp { -1/2(x-m _i ) ^t W _i ^-1 (x-m _i )}...(1) This P _i is determined for each phoneme, and the phoneme with the highest degree of similarity is determined (to processing), and the result is transferred to the phoneme recognition unit. This method is based on the idea of actively capturing changes in phonemes that are characterized by temporal changes in their spectra, such as consonants and semivowels. Figure 2 shows an example of the spectral changes of semivowels and persistent consonants. a represents a case similar to the conventional example, and the horizontal axis represents time in frames. Also, the distance between adjacent spectra on the vertical axis is shown as the Euclidean distance of the LPC cepstral coefficients. This distance curve shows the threshold value (TH) during frame number 13.
Lasts at a small value of 130ms for semi-vowels and persistent sounds.
It shows that the period of time lasts for a long time. However, in the conventional example, only the five frames indicated by circles are used because the amount of calculation required for similarity calculation is enormous. For this reason, the characteristics of semivowels and persistent sounds cannot be fully captured, and the accuracy of discrimination is low. In order to overcome this drawback, it is conceivable to use 13 frames as shown in FIG. 2b. Figure 2b
The case where the number of frames 13 is used for semivowels and persistent consonants, which is the same as in Figure 2a, is indicated by a circle. In this case, it is possible to sufficiently capture the characteristics of semi-vowels and persistent consonants, but the disadvantage is that it requires a huge amount of calculation to calculate the similarity, making the device expensive. Purpose of the Invention The present invention solves the above-mentioned drawbacks, and in automatic speech recognition, discriminates phonemes or syllables with high discrimination accuracy.
It is an object of the present invention to provide a speech recognition method that can be realized with a small amount of calculation. Structure of the Invention The present invention is intended to achieve the above-mentioned object. A standard pattern created from the voices of multiple speakers is prepared in advance, and the input unknown voice is divided into n consecutive fixed time intervals (frames). The audio is analyzed for each frame to obtain spectral information, and from the n frames, parts where the spectral change between frames does not exceed a threshold are thinned out at equal intervals. X frames (n>
Provided is a speech recognition method characterized in that phonemes or syllables are determined by calculating the similarity between the spectral information of the X frames and the standard pattern using a statistical distance scale. It is. DESCRIPTION OF EMBODIMENTS Examples of the present invention will be described below with reference to the drawings. FIG. 3 is a block diagram showing an embodiment of a device embodying the speech recognition method of the present invention. In the figure, reference numeral 1 denotes an acoustic analysis section, which analyzes audio input through a microphone or the like. As an analysis method, linear predictive analysis is performed to obtain LPC cepstra coefficients every frame period (about 10 mS). 2 is the phoneme discriminator, and the LPC obtained by the acoustic analysis unit 1
Phoneme discrimination is performed for each frame using cepstral coefficients. Reference numeral 3 denotes a standard pattern storage unit, which stores standard patterns obtained for each phoneme from the voices of multiple speakers in advance. Reference numeral 4 denotes a segmentation unit, which 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. Reference numeral 5 denotes a phoneme recognition unit, which determines what phoneme each phoneme section is based on the results of the segmentation unit 4 and the phoneme discrimination unit 2. As a result, a series of phonemes is completed. 6 is a word recognition unit that compares the 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. 7 is the word dictionary mentioned above. Next, a method for recognizing semi-vowels and persistent sounds will be explained in more detail using the flowchart of FIG. 4 as an example. This method is effective not only for semivowels and persistent consonants, but also for phonemes whose spectrum changes slowly over time, such as vowels, nasals, and fricative consonants. Statistical distance measures include Bayesian judgment, Mahalanobis distance, etc., and in this embodiment, the Mahalanobis distance will be used. Also, a case will be described in which LPC cepstral coefficients are used for spectrum information. The distinction between five vowels, semi-vowels, and persistent consonants in the word sounds of many speakers is determined in advance. The procedure for creating a standard pattern using this voice will be explained. The LPC cepstral coefficients up to the Nth order are determined for each of the n consecutive frames from the beginning of each phoneme, and up to the mth order (N≧m), that is, C'(C' ₁ ,
C′ ₂ , ..., C′m). next consecutive n
frame, so as to include at least non-consecutive frames (in this embodiment, at one frame interval)
X frames (X<n) are extracted and X C's are arranged to create C (C' ₁ , C' ₂ , . . . , C' _X ).
C is composed of M (M=X×m) cepstral coefficients. That is, C (C ₁ , C ₂ , . . . _CM ). By this C, the average value m _i (i is the phoneme name) for each phoneme and the covariance matrix W common to all target phonemes
seek. Let the inverse matrix be W ^-1 , and its (j, j′)
When the element is 〓jj′, the weighting coefficient a _ij of phoneme i with respect to C _j is obtained as follows: a _ij =2 _M 〓 ^j ′ ⁼¹ 〓jj′m _ij ……(2). Also, the average distance d _i for phoneme i is calculated as d _i =m _i ^t W ^-1 m _i (3). These a _ij and d _i are stored as standard patterns in the standard pattern storage section 3 shown in FIG. Next, the unknown audio input to the acoustic analysis unit 1 shown in FIG.
The N-th LPC cepstrum coefficients are determined for each frame, and among them, the LPC cepstrum coefficients up to the m-th order (N≧m) shown in Figure 4 Processing
Extract X′ ₁ (X′ ₁ , X′ ₂ , ..., X′m). Next, from the n consecutive frames, extract X frames so as to include at least non-consecutive frames (for example, at one frame interval in this embodiment),
As shown in the processing diagram, X frames X' ₁ , X' ₃ ,...
…, X′ _X is calculated. Furthermore, using the results of processing 1 and processing 2, as shown in processing ₂ , X× _m =M-dimensional vector
..., x _M ). Using this x, the similarity l _i is determined using the standard pattern stored in the standard pattern storage section 3 using the following equation. l _i = _M 〓 ^j=1 a _ij J ^-d i...(4) Obtain this l _i for each frame of the input audio (processing module), and remove the phoneme with the maximum similarity (processing module). ) Transfer to the phoneme recognition unit 5. The phoneme recognition unit 5 combines this result with the result of the segmentation unit 4 to create a time series of phonemes, and sends it to the word recognition unit 6. The word recognition unit 6 collates a word dictionary 7 in which phonemes are written in chronological order in advance, and outputs the word name with the highest degree of similarity as a recognition result. A specific example is shown in FIG. Figure 5a shows examples of spectral changes in semivowels and obdurate sounds. The horizontal axis shows time in frames. The spectral distance between adjacent frames on the vertical axis
It is expressed as the Euclidean distance of the LPC ceptrum coefficients.
This distance curve does not exceed the threshold value (indicated by TH) in 13 frames, indicating that the time change of the spectrum is slow. Here, the threshold value is set to a value that yields the maximum recognition rate of the target phoneme (here, a semivowel, a persistent consonant). Therefore, it is not necessarily necessary to use all frames to capture temporal changes in the spectrum. In this embodiment, a total of seven frames are used, which are thinned out every other frame and indicated by a circle. In this case, the LPC required for phoneme discrimination
The vector of cepstral coefficients is 6×7=42. Table 1 shows a comparison of the amount of calculation for product and sum per standard pattern. The conventional method shown in FIG. 2b and the present embodiment are shown separately for Bayesian determination using equation (1) and Mahalanobis distance using equation (4).

[Table] As can be seen from Table 1, this embodiment can significantly reduce the Bayesian judgment by about 30% compared to the conventional method. Furthermore, the Mahalanobis distance can be reduced to half of the conventional method. Furthermore, Table 2 shows the results of a comparison of the discrimination accuracy of semivowels and persistent consonants between the conventional method shown in FIG. 2b and the present embodiment shown in FIG. 5a.

[Table] That is, compared to the conventional method, this embodiment improves both the recognition rate and the standard deviation representing variation.
The reason for this is thought to be that by thinning out the used frames to get a broader view of the gradual temporal changes in the spectrum, it is possible to efficiently capture the characteristics. It is also believed that by removing redundant spectrum information, it is possible to reduce variations due to variable factors such as speakers and contexts. The present invention is characterized in that spectral information of a plurality of discontinuous frames is used to discriminate phonemes or syllables. Taking the case of 13 frames as an example, depending on the type of phoneme or syllable, the spectral information shown in FIG. b, c
The method also applies. In other words, the curve shown in Figure 5b represents the adjacent spectral distance of the consonant /s/.
Expressed using LPC cepstral coefficients,
Frames 1, 2, and 3 and frames 12 and 13 that exceed the threshold (TH) are used continuously, and sections that do not exceed the threshold are thinned out at equal intervals and used. This method is effective for identifying phonemes and syllables, such as consonants /s/ and /h/, in which there is a difference in the spectrum of the movement of the boundary and the frictional part. In addition, the curve shown in Figure 5c represents the adjacent spectral distance of the consonant /z/ using LPC cepstral coefficients, and frames 1, 2, 3, and 4 that exceed the threshold (TH) are used consecutively. However, the intervals that do not exceed the threshold are thinned out at equal intervals. This method is effective for identifying phonemes and syllables such as consonants /z/, /c/, and /k/, in which there is a difference in the spectrum of the movement of the plosive part and the spectrum of the frictional part. In the above embodiment, a case has been described in which continuous frames are used for the region exceeding the threshold value, but the frames do not necessarily have to be continuous. In addition, the spectral information of the present invention can be obtained by any of linear predictive analysis, band filter group analysis, and fast Fourier transform (FFT) analysis. Furthermore, the similarity calculation of the present invention is preferably performed using a statistical distance measure, and as the statistical distance measure, distance based on Bayesian judgment, Mahalanobis distance, linear discriminant function, etc. are more suitable. Effects of the Invention In summary, the present invention divides audio into n consecutive frames, analyzes the audio for each frame to obtain spectral information, and calculates the spectral displacement between frames from among the n frames. For the part where does not exceed the threshold, extract it at equal intervals to obtain X frames (n>X),
The present invention provides a speech recognition method characterized by discriminating phonemes or syllables by calculating the degree of similarity between the spectral information of these X frames and a standard pattern using a statistical distance measure. By efficiently capturing the characteristics of temporal changes in the spectra of phonemes or syllables, it is possible to obtain high recognition performance with little variation. The amount of calculation required to discriminate phonemes or syllables can be reduced to 1/2 to 1/3 of that of the conventional method, and the cost of the device can be reduced. It has the following advantages.

[Brief explanation of drawings]

FIG. 1 is a flowchart explaining phoneme discrimination in a conventional speech recognition method, FIG. 2 is a diagram explaining a conventional frame extraction method, and FIG. 3 is a diagram of a speech recognition device embodying the speech recognition method of the present invention. FIG. 4 is a block diagram showing one embodiment of the present invention, FIG. 4 is a flowchart illustrating phoneme discrimination in a speech recognition method according to an embodiment of the present invention, and FIG. 5 is a diagram illustrating a frame extraction method according to the present invention. 1... Acoustic analysis section, 2... Phoneme discrimination section, 3...
Standard pattern storage unit, 4... Segmentation unit, 5... Phoneme recognition unit, 6... Word recognition unit, 7...
...word dictionary.

Claims (1)

[Claims] 1. Divide audio into n consecutive fixed time intervals (frames), analyze the audio for each frame to obtain spectral information, and divide x frames from the n frames. (X<n) is extracted by thinning out at equal intervals for portions where the temporal change rate of the spectrum does not exceed the threshold, and the spectral information of the X frames thus extracted is extracted. A speech recognition method characterized in that phonemes or syllables are discriminated by calculating the degree of similarity between a standard pattern created in advance from the voices of multiple speakers using a statistical distance measure. 2. The speech recognition method according to claim 1, wherein the spectral information is obtained by any one of linear predictive analysis, band filter group, and fast Fourier transform analysis. 3. The speech recognition method according to claim 1, wherein the statistical distance measure is any one of a Bayesian distance, a Mahalanobis distance, and a linear discriminant function.