JPH0552512B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD The present invention relates to a speech recognition device intended for unspecified speakers. BACKGROUND TECHNOLOGY AND PROBLEMS In speech recognition, methods based on word recognition for specific speakers have already been put into practical use. This involves having a specific speaker pronounce all the words to be recognized, and then detecting and storing (registering) the acoustic parameters using a bandpass filter bank, etc.
I'll keep it. Then, when a specific speaker utters a utterance, its acoustic parameters are detected and compared with the acoustic parameters of each registered word, and when these match, the word is recognized. In such a device, if the time axis of the speaker's utterance is different from the time of registration, the time axis of the speaker's utterance is different from the time of registration,
The time series of acoustic parameters extracted every ~20 msec) is expanded or contracted to align the time axis. This makes it possible to cope with variations in speaking speed. However, with this device, the entire acoustic parameters of every word to be recognized must be registered and stored in advance.
Requires huge storage capacity and calculations. For this reason, there was a limit to the number of words that could be recognized. On the other hand, phonology (in Japanese, A, I, U, E, O, K, S, T when written in Roman sentences)
) or syllable units (KA, KI, KU, etc.) have been proposed. However, in this case, even though it is easy to recognize phonemes with quasi-stationary parts such as vowels, phonemes with very short phonological features such as plosives (K, T, P, etc.) can be recognized using only acoustic parameters. It is extremely difficult to specify one phoneme. Furthermore, when recognizing unspecified speakers,
There is a large variance in acoustic parameters due to individual differences, and recognition cannot be achieved only by matching the time axis as described above. Therefore, for example, methods have been proposed such as registering multiple acoustic parameters for one word and recognizing approximate acoustic parameters, or converting the entire word into fixed-dimensional parameters and discriminating using a discrimination function. Either method requires a huge amount of storage capacity, a large amount of calculation, and the number of words to be recognized becomes extremely small. By the way, when observing the phenomenon of phoneme production, phonemes such as vowels and fricatives (S, H, etc.) can be elongated and uttered. For example, when considering the utterance of "yes", the phoneme is as shown in Figure 1A.
Changes to "silence → H → A → I → silence". In response, the same "yes" can be uttered as shown in FIG. 1B. Here, the lengths of the quasi-stationary portions of H, A, and I change with each utterance, which causes fluctuations in the time axis. However, in this case, it has been found that there is relatively little variation in the time axis in the transitional part between each phoneme (indicated by diagonal lines). Therefore, the inventor of the present application first focused on this point and proposed the following device. In FIG. 2, an audio signal supplied to a microphone 1 is supplied to an AD conversion circuit 4 through a microphone amplifier 2 and a low-pass filter 3 of 5.5 kHz or less. Also, 12.5kHz from clock generator 5
A sampling clock (at intervals of 80 .mu.sec) is supplied to the AD conversion circuit 4, and each audio signal is converted into a digital signal of a predetermined number of bits (=1 word) at this timing. This digital signal is supplied to bandpass filters 6 ₁ , 6 _{2 .} . . 6 ₃₀ for frequency analysis, and is divided into, for example, 30 bands according to a frequency mel scale that matches human auditory characteristics. The divided signals of each band are supplied to emphasis circuits 7 ₁ , 7 _{2 .} . . 7 ₃₀ to perform high frequency enhancement in accordance with human auditory characteristics. This signal is the absolute value circuit 8 ₁ , 8 ₂ ... 8 ₃₀
is supplied to the average value circuits 9 ₁ , 9 ₂ to make it unipolar.
...9 ₃₀ is supplied and the envelope of the signal is extracted. These signals are supplied to logarithm circuits 10 ₁ , 10 ₂ . . . 10 ₃₀ and converted into logarithmic values of each signal. As a result, the above-mentioned emphasis circuits 7 ₁ , 7 ₂ ... 7 ₃₀
Redundancy due to weighting etc. is eliminated. Here, for example, if the waveform function represented by n _f sampling data included in the time length of T is Un _f T _(t) ...(1), then frequency analysis is performed to calculate the logarithm. The obtained logarithmic power spectrum log | U ² n _f T _(f) | ...(2) is the spectrum parameter x _(i) (i = 0, 1...29)
It is called. This spectral parameter x _(i) is supplied to a discrete Fourier transform (DFT) circuit 11 . Here, in this DFT circuit 11, for example, if the number of divided bands is M, this M-dimensional spectral parameter x _(i) (i=0, 1...M-1) is 2M-1
Considering the real symmetric parameters of the points, 2M−2 points
Perform DFT. _{Therefore ,} ^_ _{_} ^_ _{_} ^_ _{_} =0,1,...2M-3. Furthermore, since the function that performs this DFT is considered to be an even function, W ^mi _2M-2 = cos (2π・i・m/2M−2) = cosπ・i・m/M−1, and from these, X _(n) = _2M-3 〓 ⁱ⁼⁰ x _(i) cosπ・i・m/M−1 ...(4). This DFT extracts acoustic parameters that express the envelope characteristics of the spectrum. Regarding the spectral lamb parameter x _(i) DFT'd in this way, 0 to P-1 (for example, P=
8) Take the values of the P dimension up to the next and set this as the local parameter L _(p) (p=0,1...P-1) L _(p) = _2M-3 〓 ⁱ⁼⁰ x _(i) cosπ・i・p/M−1 ...(5) Here, considering that the spectral parameters are symmetrical, and setting x _(i) = x (2M−i−2), the local parameter L _{( p)} is L _(p) = x ₍ 〓 ₎ + _M-2 〓 〓 ⁱ⁼¹ x _(i) {cosπ・i・m/M−1+cosπ・(2M−2−
i)・p/M-1}+x _(M-1) cosπ・p/M-1 However, p=0, 1...P-1. In this way a 30 word signal is compressed into P (for example 8) words. This local parameter L _(p) is supplied to the memory device 12 . This memory device 12 has a storage section of P words per row arranged in a matrix of, for example, 83 rows, in which local parameters L _(p) are sequentially stored for each dimension, and the local parameters L (p) are stored in sequence from the clock generator 5 mentioned above. A clock is supplied at 0.96 msec intervals to sequentially shift the parameters of each row horizontally. As a result, 83 points (79.68 msec) of P-dimensional local parameters L _(p) at intervals of 0.96 msec are stored in the memory device 12, and are sequentially updated to new parameters every clock. Furthermore, the audio transition point detection circuit 20 is configured as follows. In other words, the signal V _(o) corresponding to the amount of signals in each band from the average value circuits 9 ₁ to 9 ₃₀
(n = 0, 1...29) is the biased logarithmic circuit 2
1 ₁ , 21 ₂ ... 21 ₃₀ to form v' _(o) = log (V _(o) + B) ... (7). In addition, the signal V _(o) is
This signal Va is supplied to the logarithm circuit ^21x _to form V _{' a =} log( _{V a} +B)...(8) be done. These signals are then sent to the arithmetic circuit 2.
3 to form v _(o) = v′ _a −v′ _(o) ……(9). Here, by using the signal V _(o) as described above, this signal has the same degree of change for each order (n = 0, 1...29) with respect to the change from phoneme to phoneme,
It is possible to avoid variations in the amount of change depending on the type of phoneme. Further, by forming the normalized parameter v _(o) by taking a logarithm and performing an operation, fluctuations in the parameter v _(o) due to changes in the level of input audio are eliminated.
Furthermore, by adding bias B and performing calculations, if B → ∞, the parameter v _(o) → 0
As is clear from this, the sensitivity to minute components (noise, etc.) of the input voice can be lowered. This parameter v _(o) is supplied to the memory device 24 and 2w+1 (for example, 57) points are stored. This stored signal is supplied to the arithmetic circuit 25 and Yn, t=min I∈GFt {v _(o) (I)} ...(10) where GF _t = {I; −w+t≦I≦w+t} This signal and parameter v _(o) are supplied to the arithmetic circuit 26 and T _(t) = _N-1 〓 ⁿ⁼⁰ _w 〓 ^I=-w (v _(o) (I + t) - Yn, t) ...(11) is formed. This T _t is a transient point detection parameter, and is supplied to the peak discrimination circuit 27 to detect the transition point of _the phoneme of the input speech signal. Here, since the parameter T _t is defined by w points before and after point t, there is no risk of unnecessary unevenness or multipolarity. In addition, in Figure 3, for example, the utterance of "zero" is assumed to be 12-bit digital data with a sampling frequency of 12.5 kHz, point interval = 0.96 msec, number of bands N = 30, bias B = 0, and number of detection points 2W + 1 = 57 as described above. This shows the case where detection is performed. In the figure, A is the speech waveform, B is the phoneme, and C is the detection signal.
Remarkable peaks are generated at each transition of Z, Z→E, E→R, R→O, and O→silence. Here, some unevenness is formed in the silent part due to noise, but by increasing the bias B, this becomes approximately zero as shown by the broken line. This transient point detection signal T _(t) is supplied to the memory device 12, and reading from the memory device 12 is performed when the local parameter L _(p) corresponding to the timing of this detection signal is shifted to the 42nd row. .
Here, when reading out the memory device 12, signals for 83 points are read out in the horizontal direction for each dimension P. The read signal is then supplied to the DFT circuit 13. In this circuit 13, DFT is performed in the same manner as described above, and the envelope characteristics of the time-series changes in the acoustic parameters are extracted. 0 to 0 from this DFT signal
The values of the Q dimension up to the Q-1 (for example, Q=3) order are extracted. This DFT is performed for each dimension P, and the entire transition point parameter K _(p,q) of P×Q (=24) words is
(p=0,1...P-1) (q=0,1...Q-
1) is formed. Here, since K _(0,0) expresses the power of the audio waveform, q may be set to 1 to Q when p=0 for power normalization. That is, in FIG. 4, when a transient point like B is detected for an input audio signal (HAI) like A, the entire power spectrum of this signal is like C. For example, if the power spectrum at the transition point of "H→A" is like D, this signal is emphasized to become like E, and compressed by the mel scale to become like F. This signal is subjected to DFT to become something like G, and 83 points before and after are matrixed like H, and this signal is sequentially moved in the time axis t direction.
DFT is performed to form transient point parameters K _(p,q) . This transition point parameter K _{(p, q)} is supplied to the Mahalanobis distance calculation circuit 14, and the cluster coefficients from the memory device 15 are supplied to the circuit 14 to calculate the Mahalanobis distance with each cluster coefficient. Here, the cluster coefficients are obtained by extracting transient point parameters from the pronunciations of multiple speakers in the same manner as described above, classifying them according to phoneme content, and performing statistical analysis. The calculated Mahalanobis distance is then supplied to the determination circuit 16, which determines which phoneme to which phoneme the detected transition point is a transition point,
It is taken out to the output terminal 17. For example, “Yes”, “No”, “0 (zero)”
Regarding the 12 words of ~9 (Kiyuu), the voices of a large number of speakers (more than 100 people) are supplied in advance to the above-mentioned device, the transition point is detected, and the transition point parameter is extracted. This transition point parameter is classified into a table as shown in Fig. 5, and this classification (cluster)
Perform statistical analysis for each. * in the figure indicates silence. Regarding these transition point parameters, any sample R ^(a) _r,o (r=1,2...24) (a is a cluster index, for example, a=1 is *→H, a=2 is H→A) n is the speaker number), the covariance matrix A ^(a) _rs ≡E(R ^(a) _r,o − ^(a) _r )(R ^(a) _s,o − ^(a) _s
)...(12) However, ^(a) _r = E (R ^(a) _r,o E counts the ensemble average, and this inverse matrix B ^(a) _r,s = (A ^(a) _tu ) ^-1 Find _r,s ...(13) Here, the distance between any transient point parameter Kr and cluster a is the Mahalanobis distance D( _Kr ,a)≡d〓r〓s(Kr− ^(a) _r )・B ^(a) _r,
s・ (Kr− ^(a) _s ) ...(14). Therefore, by determining and storing the above-mentioned B ^(a) _r,s and R ^(a) _r in the memory device 15, the Mahalanobis distance calculation circuit 16 calculates the Mahalanobis distance between the transition point parameter of the input voice. be done. As a result, the minimum distance to each cluster and the ranking of the transition points are extracted from the circuit 14 for each transition point of the input audio. These are supplied to the determination circuit 16,
Recognition determination is made when the input voice becomes silent. For example, for each word, the word distance is determined by the average value of the square root of the minimum distance between each transition point parameter and the cluster. In addition, taking into account the dropout of some of the transition points, word distances are calculated for multiple types assuming that each word is dropped. However, if the ranking relationship of the transition points is different from the table, it will be rejected.
Then, the word with the minimum word distance is recognized and determined. Speech recognition is performed in this way, and since this device detects changes in phoneme at transitional points in speech, there is no change in the time axis, and good recognition is possible even for unspecified speakers. It can be performed. In addition, by extracting the parameters described above at the transition point, one transition point can be
It can be recognized in 24 dimensions, making recognition extremely easy and accurate. Furthermore, as a result of learning using the above-mentioned device with 120 speakers and conducting experiments on the 12 words mentioned above with speakers other than these 120, an average recognition rate of 98.2% was obtained. Furthermore, in the above example, “H → A” of “Yes” and “8”
“H→A” of “(Hachi)” can be classified into the same cluster. Therefore, if the number of phonemes of the language to be recognized is α, αC ₂ clusters are calculated in advance and the cluster coefficients are stored in the memory device 15.
It can be applied to the recognition of various words, and can easily recognize many words. By the way, in the above example, specific words such as "yes" and "no" were recognized, but it is also possible to recognize general speech, for example, on a monosyllable basis. However, in that case, the number of phonemes in human pronunciation is large, and therefore the number of clusters of transition points is 100.
~200, which is extremely large. For this reason, for example, if an attempt was made to calculate the Mahalanobis distance for all these clusters, the amount of calculation would be extremely large, making it impractical. For example, when recognizing a single syllable, when looking at the last vowel → silence, multiple transition points may occur due to fluctuations in the voice level, and the vowels in these cases may be different. In that case, it was found that the one with the smallest Mahalanobis distance was not necessarily the phoneme at that time. OBJECTS OF THE INVENTION In view of the above points, the present invention enables good speech recognition to be performed with a simple configuration. Summary of the Invention The present invention detects a transitional part between phonemes including silence, extracts a predetermined length of sound in the detected transitional part, converts it into a parameter, and uses this parameter as a basic unit of speech recognition. In the recognition method,
A speech recognition method characterized in that, when there are multiple transition points from a vowel to silence classified into different cluster coefficients, a cluster coefficient is determined based on the number of transition points classified into each cluster coefficient, According to this, good speech recognition can be performed with a simple configuration. Embodiment By the way, in the following embodiment, the following apparatus is used. That is, in FIG. 6, the emphasis circuit 7 is provided before the bandpass filters ₆₁ to ₆₃₀ .
is provided. In this emphasis circuit 7, for example, in the bands 1 to 16 on the low frequency side, the signals are supplied to the bandpass filters ₆₁ to ₆₁₆ without correction, and in the bands 17 to 30 on the high frequency side, the signals are supplied to the bandpass filters 61 to 616 by the difference. The signal is supplied through the circuit 31 to bandpass filters 6 ₁₇ to 6 ₃₀ . In this emphasis circuit 7, the differential circuit 3
_{The characteristics of _} ^_ _{_ o)} ……(16) becomes. Furthermore, the transfer function H _(z) of this circuit is |H(Z)| ² = |H(Z)・H _(z-1) | = |2−2cosωT| ...(17), as shown in Figure 7. The characteristic is that it is small in the low range and large in the high range. This transfer function becomes 1 at the point where the angular frequency ω becomes π/2. On the other hand, when dividing into 30 bands on the above-mentioned mel scale, the point where the angular frequency ω is π/2 is between the 16th and 17th bands. Therefore, as mentioned above, by making no correction in the 1st to 16th bands and using a difference in the 17th to 30th bands, it is possible to perform high-frequency enhancement that matches the human auditory characteristics as shown in Figure 8. . Further, signals from the average value circuits 9 ₁ to 9 ₃₀ of the respective bands are supplied to the noise removal circuits 32 ₁ to 32 ₃₀ . On the other hand, the signal from the AD conversion circuit 4 is supplied to the silent state detection circuit 33, and this detection signal is supplied to the removal circuits 32 ₁ to _{32 30} . Then, in the removal circuits 32 ₁ to 32 ₃₀ , the signal (noise) in a silent state is measured, and the average value (or peak value or the value obtained by calculating these) is set as the threshold level N, and the input signal is When x is smaller than this level N, a signal of 0 is output, and when x is larger, a signal of (x-N) is output. This signal is supplied to logarithmic circuits 10 ₁ to 10 ₃₀ . That is, in the noise removal circuits 32 ₁ to 32 ₃₀ , when a signal as shown in FIG. 9A is supplied to the removal circuit of one band, the detection circuit 33 detects a silent part, and the signal of this part A signal as shown in FIG. 9B is output depending on the threshold level N, which is, for example, an average value. In this case, the noise level is measured for each band, and noise removal is performed according to the frequency characteristics of the noise. The rest of the structure is the same as in FIG. According to this device, it is possible to perform good emphasis matching the human auditory characteristics using only a simple differential circuit without using a multiplier. Furthermore, the amount of calculation can be reduced when processing is performed using software. Furthermore, noise removal can be performed according to the frequency characteristics of the noise, and the accuracy of parameters is greatly improved. In this device, the distance calculation circuit 14 and the determination circuit 16 are configured as follows. That is, in FIG. 10, the signal from the DFT circuit 13 is supplied to the first distance calculation circuit 41, and the distance from the cluster coefficient from the memory device 51 is calculated. Here, the memory device 51 stores information such as [*→ (indicates a voiced sound),” “→ (indicates a vowel),” “→
Three types of cluster coefficients, ``*'', are written. A monosyllable is formed by these three transition points. Furthermore, the calculated distance is supplied to the first determination circuit 61, and the input transient point parameters are classified into the three clusters described above. Among the classified parameters, the "→*" parameter is supplied to the second distance calculation circuit 42, and the distance from the cluster coefficient from the memory device 52 is calculated. Here, the memory device 52 stores “A→*” and “I→
*” “U→*” “E→*” “O→*” “→*” (
6 types of cluster coefficients are written. Furthermore, the calculated distance is supplied to a second determination circuit 62, which determines which of the six clusters the input parameter corresponds to. Furthermore, this determination result is supplied to the processing circuit 71. Here, in this circuit 71, comprehensive determination of vowels is performed. That is, at the transition point of "→*", a plurality of transition points may be detected due to noise components such as so-called bulges, and in that case, there is a possibility that parameters close to other clusters may appear by chance. Therefore, in the processing circuit 71, the calculated distance and the number are comprehensively determined. That is, when a determination result and distance as shown in B are calculated by detecting a transient point as shown in FIG. 11A, for example, the shortest distance is "U".
However, in this case, the number determined is "A". As a result of conducting experiments and simulations on such cases, it was found that in such cases, it is generally correct to have more. Therefore, in this processing circuit 71, for example, determination is made by majority vote on the transition point parameters. Note that if the number is the same by majority vote or if the distances are extremely different, these distances may be taken into consideration. In this way, the final vowel is determined. In addition, the transition point parameters of “*→” and “→” classified by the determination circuit 61 are
are supplied to distance calculation circuits 43 and 44, respectively, and the distances from the cluster coefficients from memory devices 53 and 54 are calculated. First, cluster coefficients as shown in the table below are written in the memory device 53, classified by final vowel.

【table】

[Table] Here, for example, the clusters classified as the final vowel "A" are the 10 A-stages of the 50-tone table, the 5 voiced and semi-voiced
"＊→" for 27 sounds, 11 sounds, and 27 buzz sounds.
There are a total of 32 sounds, including 5 plosives that are difficult to judge, such as "→". In addition, "I" has a total of 15 characters, excluding the "Y", "Wa", "DA", and "Su" sounds from "A". Below, "U", "E", and "O" are also composed of 30 clusters, 17 clusters, and 31 clusters, respectively, according to their pronunciation characteristics. Note that ``'' is included in ``''. Further, cluster coefficients as shown in the table below are written in the memory device 54, classified by final vowel.

【table】

[Table] Here, as in the case of the memory device 53 described above, 26 "A"s, 26 "A"s,
It is classified and written into clusters of 12 "I", 25 "U", 13 "E", and 25 "O". Note that the syllables may be integrated into "Y→A,""Y→U," and "Y→O." Furthermore, the same plosive sounds as in the memory device 53 are repeatedly provided. Then, in accordance with the determination output of the final vowel from the processing circuit 71 described above, only the corresponding vowel portion of each memory device 53, 54 is supplied to the calculation circuits 43, 44, and distance calculation is performed. The distances are further calculated and the third and fourth distances are respectively calculated.
The input parameters are supplied to determination circuits 63 and 64, which determine which of the respective clusters the input parameters correspond to. These determination results and the determination result from the determination circuit 62 are supplied to a word/single syllable determination circuit 81 to identify words/single syllables of the input speech. Speech recognition is thus performed in this device.According to this device, first, the transition points are classified into three types, and then the final vowel is determined.
In general, vowel detection is easy, and the first three classifications and vowel determinations are performed when the number of clusters is 3 and 6.
Therefore, extremely accurate judgment can be made by increasing the number of dimensions of the parameters. Further, when a plurality of final vowels are detected, the accuracy of the determination can be further improved by comprehensively determining them based on distance and number. By using this determined final vowel to limit the cluster of previous transient point detections, the amount of calculation for these distances can be reduced, making it easier to implement and improving accuracy. It can also be increased. Effects of the Invention According to the present invention, it has become possible to perform good speech recognition with a simple configuration.

[Brief explanation of the drawing]

Figure 1 is a diagram for explaining audio, Figures 2 to 5
The figures are for explaining the conventional device, Figures 6 to 9.
The figures are diagrams for explaining the present invention, FIG. 10 is a system diagram of an example of the present invention, and FIG. 11 is a diagram for explaining the same. 1 is a microphone, 3 is a low pass filter,
4 is an AD conversion circuit, 5 is a clock generator, 6 is a band pass filter, 7 is an emphasis circuit, 8 is an absolute value circuit, 9 is an average value circuit, 10 is a logarithmic circuit,
11, 13 are discrete Fourier transform circuits, 12, 1
5, 51-54 are memory devices, 14, 41-44
is Mahalanobis distance calculation circuit, 16, 61-64
is a determination circuit, 17 is an output terminal, 20 is a transient point detection circuit, 31 is a difference circuit, 32 is a noise removal circuit,
33 is a silent part detection circuit, 71 is a processing circuit, and 81 is a word/single syllable determination circuit.

Claims (1)

[Claims] 1. A speech in which a transitional part between phonemes including silence is detected, a predetermined length of speech in the detected transient part is extracted and converted into a parameter, and this parameter is used as a basic recognition unit. In the recognition method, when there are multiple transition points from a vowel to silence classified into different cluster coefficients, the cluster coefficient is determined based on the number of transition points classified into each cluster coefficient. Recognition method.