JPS58145999A - Recognition of voice - 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 The present invention relates to speech recognition and proposes an f# method.

In speech recognition, methods based on word recognition for specific speakers have already been put into practical use. This involves having a specific speaker say all the words to be recognized, and detecting and storing (registering) the acoustic parameters using a bandpass filter bank or the like. Then, when a specific speaker utters a utterance, its acoustic parameters are detected and compared with the sound parameters of each registered word, and when these match, it is determined that the word is the same.

In addition, 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 for each interval is expanded or contracted to align the 1-time axis. This makes it possible to cope with fluctuations in speaking speed.

However, in the case of the above device, all single INs to be recognized
The entire acoustic parameters of the word must be stored in advance, which requires a huge amount of storage capacity and calculations. For this reason, there was a limit to the number of words that could be recognized.

On the other hand, recognition in terms of phoneme (in Japanese candy, when written in the ro! character, A, I, U, E, 0., 8.T, etc.) or sound @ (KA, KI, K1, etc.) It is proposed to do so. However, in this case, K, even though it is easy to recognize the phoneme with the quasi-stationary part of the vowel cape,
It is extremely difficult to identify plosives (K, T, P fathom), which have very short phonetic features, as a single phoneme using only acoustic parameters.

Conventionally, the discretely pronounced sounds for each syllable are recorded, and the discretely pronounced sounds are converted into the word wtIm! Similarly, the ljg iron was performed in a time-aligned manner, and could only be used for limited purposes to produce special vocalizations.

Furthermore, when an unspecified speaker is targeted for s1wIt, there is a large variance in the acoustic parameters due to individual differences, and recognition cannot be performed only by matching the time axis as described above. Therefore, for example, there is a method of registering multiple acoustic parameters for one word and elaborating the approximate Aotatami parameters.
Methods have been proposed in which the entire word is converted into identification-dimensional parameters and discriminated using a discrimination function, but these methods require a huge amount of storage capacity, a large amount of calculation, and are difficult to implement in i&ili.
The number of woven candy will be significantly reduced.

In view of these points, the present invention proposes a novel speech recognition method that allows speech recognition to be easily and reliably performed even for unspecified speakers. An embodiment of the present invention will be described below with reference to the drawings.

By the way, when we observe the phenomenon of phonological vocalization, we find that vowels and fricatives (
For example, when considering the pronunciation of ``1'', this phoneme is pronounced as K, ``silence → H → A → Changes from “I → Silence”. In response, make the same 1 yes 10 utterance as the first
It can also be done as shown in Figure B. Here, the lengths of the quasi-fixed f parts of H, A, and I change depending on the utterance, and this causes a fluctuation in the time axis. However, in this case, it has been found that the transition part between each phoneme (indicated by one line) has relatively little variation in the time axis.

The inventor of the present application has focused on this point.

In FIG. 2, an audio signal supplied to a microphone amplifier (1) is supplied to an A/D conversion circuit (4) through a microphone amplifier (2) and a low-pass filter (3) of 5.5 kHz or less. IL5kHz from Matari error generator (5)
1 (80 tones (6) intervals) Nosu/Prindak Yotsuku is AD
The audio signals are supplied to a conversion circuit (4), and each audio signal is converted into a digital signal of a predetermined number of bits (=1 word) by this tie writing. This converted audio signal is supplied to a 5×64 word register (6)K. In addition, a frame clock with a 5.12-frame interval from the clock generator (5) is supplied to a quinary counter (7), and this count value is supplied to a register (6) to shift the audio signal by 64 words. The shifted 4×64 word signal is taken from register (6).

4 x 64 = taken from this register (6)
254! The word signal undergoes a fast Fourier transform ()'FT
) Circuit (8) K is supplied. Here, this FF7 circuit (8
), the example is n1 times included in the time length of T/
The waveform function represented by the pulling data is U n fT(t) −(
When It is assumed, this Fourier transform becomes 〒 =U ln fT(f)
=-(21 signals are obtained.

Further, the signal from this FF7 circuit (8) is supplied to the worst vector detection circuit (9), and 10"l = U"s n fT(f) +uinft
(t) The power spectrum signal of H4+44H(a) is extracted. Here, since the Fourier transformed signal is symmetrical on the frequency axis, half of the IIf data extracted by the Fourier transform is redundant data. Then half of the data is removed and inf pieces of data are extracted. That is, the above-mentioned 0FFT enclosure (8)
The 2s6 word signal supplied to is converted and the H8 word /(worth vector signal is taken out.

This power spectrum signal is supplied to an emphasis circuit and weighted for auditory correction.

Here, as the weighting, for example, correction is performed to enhance high frequency components.

This weighted signal is supplied to the band division filter, and is divided into, for example, 32 bands according to a frequency mel scale matched to auditory characteristics. If the dividing point is different from the dividing point of the Worth vector, the signal is divided into each band proportionally.
Signals corresponding to the amount of signals in each band are extracted. As a result, the above-mentioned 128-word power spectrum signal is compressed into 32 words that preserve the blue-tatami characteristics.

This signal is supplied to the logarithm circuit a, and is converted into a logarithmic value of each signal. This eliminates the redundancy due to weighting in the above-mentioned emphasis circuit. Now this logarithmic power spectrum tog I l4t(f) I
・・・－・・・(4)
Spectral parameter
1).

This spectral parameter x(1) is supplied to a discrete Fourier transform (DFT) circuit (13, where this DF
In T-connection QI, for example, if the number of divided bands is M, then this M-dimensional spectral parameter X mountain (i = 0
, 1 . . . M-1)- is regarded as a real symmetric parameter of 2M points, and CDFT is performed. Therefore, 1M = 0, 1...2M-1. Furthermore, since the function that performs this DFT is considered to have an even number of strokes, goi*w* = 顛□-, and from these. This DFT extracts acoustic parameters representing the envelope characteristics of the spectrum.

This is 5KL, the smooth and smooth fi that was DFT
X<i>K tweet, 0-P-1 (r=8 if fi)
If we take the values of the V dimension up to the next and set this as the local parameter L (p) (p=0.1...P-1)...(7), where the spectral parameters are symmetrical. Considering a certain thing, X (Todoroki) ” X (xli-i-x)
-----If (8) is set, the local parameter L
(p) is...(9) However, p is 0,1...P-1. This 5KL, 32-word signal is divided into r (for example, 8) words.

This local parameter L (p) is supplied to the memory device I. This memory device I has a storage section of r words per row arranged in a matrix of 16 rows, for example, and stores the local parameter L(p) or each dimension in KJI 11th order, and also stores the above-mentioned clock generator (5 ) from 5.12
A frame clock at intervals of one vehicle is supplied, and the parameters of each row are sequentially shifted in the horizontal direction.

As a result, [Memory device I has an r space of 5.12m5c apart.
The local parameter of the dimension, L(p), is 167 remes (
81,92m5ec) is stored and updated to new parameters sequentially at every frame clock.Furthermore, the weighted signal from the emphasis 1 path C11) is supplied to the band division path C11), and the weighted signal is stored according to the mel scale in the same way as described above. The signal is divided into N (for example, 20) bands, and the signal is divided into N (for example, 20) bands according to the amount of signals in each band.
=o, 1...N-1) are extracted. This signal is supplied to the biased logarithm circuit 5 (to) k and V'(n
) = ”g(V(n) + B)
...=H is formed. Also, the signal VOI) is supplied to the accumulator circuit (2) and v-=jog(V10B)
−...Qυ is formed. These signals are then supplied to an arithmetic circuit (product) to form V(n)=ma-v color)-a side.

Here, by using the signal V(n) as described above,
This signal corresponds to each order (n=o
, 1 .

In addition, the logarithm is taken and the calculation is performed to normalize the parameter V(II
) eliminates fluctuations in the parameter V(n) due to changes in the level of the input audio. By further adding bias B and performing the calculation, it becomes [3
-* If oo, the buff meter v01) → 0. As is clear from this, the minute components of the input voice (noise)
sensitivity to can be lowered.

This parameter V(A'') is supplied to the memory device (c) and 'C2W+1 (for example, 9) frames are stored.

This stored signal is supplied to the arithmetic circuit (e) and Yn
, t-1,:'ffFN(vQl)(■))...
・(B) However, GFN = (I; - fact + weight ≦■≦w-) -
1) is formed, and this signal and the buff meter y<t> are supplied to the arithmetic circuit K to form a. This T(t) is the transient point detection parameter, T(t) is supplied to the beater discrimination circuit (to), and the transition point of the phoneme of the input audio signal is detected.

Here, the parameter T(t) is defined by W7 frames before and after frame 5 t, so unnecessary unevenness and f
There is no risk of producing IkIL. Note that the third wJ is, for example, 5 pronunciations such as 0 zee, and the sampling frequency is ILikH.
i, l-bit digital data, perform FFT of 26 points with 5.1 half-frame period, number of bands h = 0, bias x Bm6. This shows the case where the above-mentioned detection is performed with the number of detected 7 frames being 2W+1w-1. In the figure, A is the audio waveform, B
is the phoneme, C is the detection signal, and “silence-4Z J
l" Z-4K J "g -4RJ "R→0""
A remarkable peak occurs at each transition from 0 to silence. Here, if some unevenness is formed in the silent part due to noise, this can be substantially corrected by increasing the bias B (as shown in the diagram).

This overflow point detection signal T(t) is supplied to the measuring device I, and at the time when the 11 Kalbat meter L(P), which corresponds to the tie liner of this detection signal, is shifted to the 8th row. Reading of the printing device 1 is performed. Here, when reading out the memory scissors *ai, signals corresponding to 16 frames and 15 minutes are read out in the horizontal direction for each dimension P. Then, the read signal is DFT
It is supplied to circuit Q5.

In this circuit (lstm, DFT is performed in the same manner as described above, and the envelope characteristics of the time-series changes in acoustic parameters are extracted.
=3) Extract the values of the Q dimension up to the next one.

This DFT is performed for each dimension P, and the total is PxQ (=2
4) Word transition point parameter K(p, Q) (p=o
, 1...p-1) (Q=0.1・hit・Q-1
) is formed. Here, K(0,0) is a constant, so
When p=o, Kq may be set as 1-Q.

That is, in FIG. 4, if an input audio signal like A (H
If a transient point like B is detected for AI), the entire power spectrum of this signal is like C. For example, if the power spectrum at the transition point of [H→A] is as shown in D, this signal is emphasized to become as shown in E, and compressed using the mel scale as shown in F. This signal is subjected to DFT to become something like G, 16 frames and 15 minutes before and after are matrixed to the quality of H, and this signal is sequentially DFT'd in the direction of the time axis to form a transient point parameter K (psq). .

This transient point parameter K (p, q) is supplied to the Mahalanobis distance calculation circuit QIK, and the cluster system from the memory device aη is supplied to the circuit (1G) [the Mahalanobis distance with each cluster system is calculated].

Here, the cluster system is obtained by extracting transition point parameters of the pronunciations of a plurality of speakers in the same way as described above, and classifying and statistically analyzing them according to the self-answered phonemes.

The calculated Mahanobis distance is then supplied to the determination circuit 0, which determines which phoneme to which phoneme the detected transition point is a transition point, and outputs it to the output terminal a.

For example, 1 yes "1 no" 0 (zero) 1~@9
(Kieu) For 1012 words, the voices of a large number of speakers (more than 100 Å) are supplied in advance to the above-mentioned device, a transition point is detected, and a transition point parameter is extracted. The transient point parameters are classified into a table as shown in FIG. 5, for example, and statistically analyzed into # (cluster). * in the figure indicates silence.

For these transition point parameters, any S (1) /pull is Ry, B (r := 112"=・24
) (” is a cluster index, for example -=1 is *→H, a=
2 corresponds to H→A. - is the speaker number), and the covariance matrix...ae However, R? )−E(Rshi5) E is the number of stitches of the unnumbered average, and this inverse matrix (1) (a)−1 Br, s” (At, u) r, s
--- Find Qi.

Here, the distance between an arbitrary transition point parameter and cluster 1 is determined by Mahalanobis' 1n (K·″″RRso゛ . . . Q7).

Therefore, by determining and storing the above-mentioned B, s, and l in the memory processing device Qη, the Mahalanobis distance between the transition point parameter of the input voice and the input voice is calculated by the Mahalanobis distance calculation circuit (IQ).

As a result, the minimum distance to each cluster and the ranking of the transition points are extracted from the C circuit loss Q for each transition point of the input audio. A recognition judgment is made when the

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.

? j Considering the omission of some of the transition points, [calculate the word distance for multiple types assuming that each word is omitted. 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.According to the present invention, changes in the phoneme at one point in the speech are detected, so there is no change in the time axis, and it is possible to perform good g-strings for unspecified speakers. be able to.

Furthermore, by extracting the parameters as described above at a transition point, it is possible to identify one transition point in, for example, 24 dimensions, which can be done extremely easily and accurately.

In addition, as a result of learning with the above-mentioned device using 120 speakers and conducting experiments on the above-mentioned 12 words with speakers other than the 120 speakers, an average wl recognition rate of 96.5% was obtained.

Furthermore, in the above example, 1 is 1's "H → AJ do 8 (Hachi)"
0's "H-*AJ" can be classified into the same cluster. Therefore, assuming the number of phonemes of the language to be recognized as α, 1oLC1 clusters are calculated in advance and the cluster coefficient is stored in memory. If this method is established, it can be applied to a single recognition of a seed yam, and recognition of many words can be easily performed.

[BRIEF DESCRIPTION OF THE DRAWINGS] The irises 11Q are diagrams for explaining audio atomization, FIG. 2 is a diagram showing a system diagram of an example of the present invention, and diagrams 3 to 5 are diagrams for explaining the same. (1) Hamaitalophone, +31 is low pass filter,
(4) is an AD conversion circuit, (5) is a clock generator, (6
) is a register, (7) is a kakunta, (8) is a fast Fourier transform circuit, (9) is a power spectrum detection circuit, ae
is an emphasis circuit, ae is a band division circuit, a4 is a pair I
I [circuit, 03.6 is a discrete 7-lier transform tunnel, axis,
＠ is a memory device, the axis is a Machatsunobis distance calculation circuit, the brocade is a determination circuit, 0 is an output terminal, and υ ~ (to) is a circuit for detecting a transition point. No. 5 i”1 Procedural Amendment May 25, 1939 Showa S8 1'1' - Indication of Patent Application No. !-412 2, Title of Invention Sound - Weaving Method 3, Amendment Relationship with Patent Applicant Address: 6-7-35, Kitashina-yo, Tokyo Parts-Yo-ku Name (21
8) Norio Ohga, Representative Director of Sony Corporation 6. The number of inventions will increase due to the supplement (1) In the specification, line 7119, F2M 4 points.''
2M-1 point,” he corrected. (2) Same, same] 110 lines 1 DFT & line 5”
Do a Dri for M-2 points,” he corrected. (3) Same, Doshin lines 11-14 [x(,,,)-7Gen1)/"i-O謹=S2X(i)W[f)' di......(5
) Proverb m mm0.1...-2M-IJ mmo, 1.
...2 to 3," he corrected. (4) Oka, Iris 8 Yu lines 1-2 i [W mm (traitors) 1M4 2M-yuki=biased (five'' two!-) Corrected. -1 (5) Oka, same line 114 [X(m) = "ff'L<p>-5, xO) Seagull] "" +7) rllj, line 93N2 r X(i) = X (sM-i-t)...
・(8)'' It is corrected as ``X(i) =-X (-breath-ri)'' ・(8) Same, same line j14 (9) Same, 10th line 10 aO same, Mlllkl! i line ry<th>J and r
Ycn) Correct it as J. aυ Same, same jll, line 17 ag Same, line 131111, correct the statement ``Because it is a constant'' to ``Because it expresses the power of the voice drum shape, it is used for power normalization.'' as same, irises 14116, 7, 8 lines sotL sot
``L r/j star series'' is corrected to ``Cluster coefficient''. Q4 Same, Iris line 17116 says "96.5%".
9s, zou,” he corrected. as same, same jal 1 line rcLCm pieces” rc
"About 1 IPs," he corrected. that's all

Claims (1)

[Claims] A speech method that has means for detecting a transitional part between phonemes including silence, extracts a predetermined length of speech in the detected transient part, converts it into a parameter, and uses the parameter as a basic recognition unit. .