JPS61228500A - Voice recognition - 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 (Field of Industrial Application) The present invention relates to a speech recognition method, and more particularly to a speech recognition method that performs word speech recognition using local beams.

(Prior art) Conventionally, this type of speech recognition method was published in Journal of the Institute of Electronics and Communication Engineers, J68-A (1) (January 1985) p-78-8.
There were items listed in 5. Figure 2 is a flowchart of a conventional speech recognition method using local peaks.
The input voice is frequency-analyzed at every 10 m5ec by a group of 15-channel bandpass filters (see 1 in Figure 2), and the voice spectrum is expressed logarithmically on both the amplitude and frequency axes as a method for normalizing individual differences in vocal cord sound source characteristics. Find the least squares approximation straight line (see 2 in Figure 2) and correct it by taking the difference.

However, if the slope of the least squares approximation line is positive, the difference from the average value is taken. After that, as shown in Fig. 3, for each frame (10m+sec), for each part where the amplitude is OdB or more, the channel that has the maximum value among those with amplitudes of 172 or more of each maximum value is considered as the channel with a local peak. Binarization is performed by setting the value to ``1'' and the others to @0 (see 3 in Figure 2).The number of channels of the bandpass filter is 15, but the slope of the least squares approximation line is negative at the 16th channel. When the slope is considered to be a voiced sound, a value of 1 is set, and when the slope is positive, it is considered to be an unvoiced sound and a value of "0" is set, and a slope sign is added (see 4 in Fig. 2).

The weighted average dictionary is obtained as a multivalued pattern by linearly extending and adding a plurality of binarized patterns to the longest one on the time axis (see 5 in FIG. 2).

To match a binary input pattern and a multi-value weighted average dictionary, linearly extend the time direction to the longer / pattern, perform calculations based on a certain degree of similarity, and calculate the maximum degree of similarity. The category name of the given standard tsukutan is used as the recognition result (see 6 in Fig. 2).

(Problems to be Solved by the Invention) However, the conventional methods described above function effectively in environments with a good signal-to-noise ratio, such as when using a close-talking microphone, but they do not work particularly well in environments with strong noise. There was a problem in that it was easy to pick up local noise peaks in silent parts, leading to an increase in false recognition.

The present invention eliminates the problem of conventional technology in which noise peaks are extracted as local peaks of speech, especially in silent parts, in environments with strong noise, resulting in increased erroneous recognition, and provides a speech recognition method that is highly resistant to noise. (Means for solving the problem) The speech recognition method according to the present invention first performs frequency analysis of input speech into analysis data of multiple channels for each speech frame. Further, all analysis data of all voice frames in a predetermined noise section, which is known in advance to be only noise, is averaged to calculate the noise average power. Then, for each audio frame of the input audio, all analysis data corresponding to the frame is averaged to calculate the average power. Next, the average voice power, which is the difference between the above-mentioned average power and the above-mentioned noise average power, is calculated for each voice frame. On the other hand, analysis data of input speech is spectral normalized using a least square approximation straight line of the speech spectrum in the speech frame to which the data belongs. For this spectrum-normalized analysis data, the presence or absence of local maximum values along the frequency axis, that is, local peaks, is determined using the above-mentioned least squares approximation line as a reference). Conversion to a binarized local peak pattern is performed with the case 0'' and a local peak i<? prepared in advance. The degree of similarity between multiple standard patterns represented by buttons and the local peak pattern of the input voice is calculated for each normalized voice frame between the standard pattern and the input voice, and □The above-mentioned voice average power is calculated based on this similarity. The final similarity is calculated by weighting and accumulating over all the normalized audio frames. As a result, the standard No.4-tan category name corresponding to the final similarity that gives the maximum value is determined to be the recognition result.

(Function) The present invention calculates the average voice power, which is the difference between the average power of the frequency analysis data of the input voice and the noise average power of a predetermined noise section, and calculates the local peak pattern of each standard/mother button and the input voice. Since the final similarity is calculated by weighting the similarity with this voice average power, it is possible to compensate for the incompleteness of local peak extraction for noisy voices, especially low-level parts of the voice. can.

(Example) FIG. 1 is a flowchart of the speech recognition method according to the present invention, in which the input speech signal is frequency-analyzed using Stella fxr.
The frequency is analyzed every 10m5ec (frame period),
Logarithmic values as analysis data of multiple channels are output for each audio frame.

Next, based on the above logarithmic data of the noise interval given in advance,
The noise average power calculation Stella 7p12 calculates the noise average power by averaging all the logarithmic values of all the speech frames within the noise interval.

On the other hand, the voice average power calculation Stella f13 calculates the voice average power using the difference between the average power found from the frequency-analyzed logarithmic value of each voice frame of the input voice and the noise average power.

Further, the logarithmic value of the frequency analysis data of the input speech is spectral normalized using the least square approximation straight line of the speech quictor in the speech frame to which the logarithmic value belongs (step 14). Next, determining the presence or absence of a local maximum value, that is, a local peak, along the frequency axis with respect to the spectrum-normalized logarithm value using the least squares approximation straight line as a reference;
The input local peak pattern is converted into a binary input local peak pattern in which "1" indicates the presence of a local peak and "O" indicates no local peak.

The standard pattern 16 includes a plurality of binary local peaks A'
It is created by linearly expanding and contracting the tan to the average utterance length for each category and adding the results.

The matching Stella f17 linearly expands and contracts the binary input local peak/4tan to the frame length of each standard pattern, calculates the degree of similarity with each standard, and calculates the similarity with the average voice power calculation Stella f13. A certain speech average power is weighted to the similarity, and the weighted similarity is accumulated over all speech frames to obtain the final similarity. and the standard/J that gives the maximum final similarity? Output the tongue category name as the recognition result.

Next, an embodiment of a speech recognition apparatus using the speech recognition method of the present invention will be described. FIG. 4 is a diagram showing the overall circuit configuration of an embodiment of the speech recognition apparatus according to the present invention.

In this embodiment, as shown in FIG. 4, 101 is an audio input terminal, 102 is a frequency analysis section, 103 is a hash vector normalization #ZOS is an average power memory 1106 is a noise power calculation section, and 107 is a speech section detection section. , 108 is a voice average power calculation unit, 109 is a local peak feature extraction unit, and 110 is a standard iJ? Tan memory, 111 is a linear matching section, 11
2 is a determination unit, and 113 is an output terminal for recognition results.

The operation will be explained below.

First, an audio signal converted into a stain signal by a microphone or the like is input to the audio input terminal 101, and is then input to the frequency analysis section 10.
2, the frequency spectrum is extracted. The circuit configuration of this frequency analysis section 102 is shown in FIG.

In FIG. 5, 201 is a preamplifier (Amp);
2 is a group of band pass filters (BPF-1t...).

BPF-N), 2os is a full wave rectifier group (DET-1
#...

DET-N), 204 is a low-pass filter group (LPF
-1°...#LPF-N), 2os is an analog multiplexer, 206 is an AD converter, and 207 is a ROM for logarithmic conversion. Band iJ? The filter groups are arranged from 1 to N at logarithmically equal center frequencies, starting from the lowest N frequencies, and are hereinafter referred to as a first channel, . . . , an Nth channel. In this embodiment, N=22.

The output of the low/gus filter group 204 is switched by the multiplexer 205, converted to a digital quantity by the AD converter 206 at a sampling period of about 10 ms, converted to a logarithmic value by the logarithmic conversion ROM 207, and then It is sent to the normalization unit 103.

In this embodiment, an analog filter is used in the configuration of the frequency analysis section 102, but the configuration is essentially the same even if a digital filter bank is used.

Logarithmic conversion ROM 207 K outputs the audio spectrum data for N channels of the 2nd frame to
(K) ,...,X note. Spectrum normalization section 1
In step 03, first, the average value or ■ is calculated for each frame based on the following equation (1).

Hereinafter, (8) will be referred to as the "average power" in the frame.

The average power theory is stored in the average power memory 105 and simultaneously sent to the noise power calculation unit 106.

In the noise power calculation unit 106, according to an instruction input from the outside, the section [KE, which is known to be only noise] is detected in advance.

Calculating the average noise power in [KE2] @The instructions for the noise-only section are known in advance by asking the speaker to stop speaking.

To measure the noise power, the average noise power N in the noise interval (KEl, Kl□) is obtained as shown in the following equation (2).

In this example, K = 20 to 30 (frames)
It is said that As the next step, the voice section detection unit 10
In step 7, a voice section is detected.

Speech interval detection can be performed, for example, by using two thresholds E for the average power.
, , K2(El)K2). In other words, as shown in Figure 6, the average power exceeds K2, and then E
When exceeding E, which becomes smaller than 2, let the point beyond K2 be the starting point. Similarly, reverse the time axis and find 2 at the end, [K1. 2] is the voice section.

Threshold E1. E2 is the noise average power N, where Lf, K
2 is an arbitrary constant, L,>K2>0.

It is closed.

In addition, the spectrum normalization unit 103 generates audio spectrum data X1(6) (K==1..., 2:i=1
.

Y amount α0=A(K) ・i+BσQ
(4) That is, the difference between the audio spectrum data Xi■ and the least squares approximation straight line is calculated using the following equation (5) to obtain the spectrum normalized data z i Cyst.

Zi([= Xl(IFO(A(K) ・i + B(
K)) (K==, #..., 2; i=1.-,2
)・-(5) ACKJ, B in the least squares approximation straight line of the above equation (4)
(6) is the next (6)? It is given by equation (7).

(7) The calculation of the above equation (5) is performed when AI) is positive, that is, in most cases, even for unvoiced sounds. In the time direction, equation (5) is calculated for all frames K before the voice section is detected. Also, the average power of the voice section
) + Kv [K1 y K2 :) and noise average power N
, X(8), which is called the audio average power, is calculated from .

(K==4. . . , 2) (8) The calculation of the above equation (8) is performed by the audio average power calculation unit 108. In other words, as X'(6), mix (0-X(EI
Np).

Next, the local peak feature extraction unit 109 extracts spectrum normalized data ZtCKL (
K=, t·s K2; t==t pp N) are sequentially processed for each frame.

The detailed configuration of the local peak feature extraction section 109 is shown in FIG.

In FIG. 7, 40 is a data storage memory, 402 is a peak extractor, and 403 is a local peak pattern storage memory.

The data storage memory 401 stores spectral normalized data Zi(6) from 1 at the beginning to 2 at the end (K=1°...
, K2; i = 1, 2 *..., N) are stored, and the peak extraction unit 402 extracts spectrum normalized data Zi(6) (K = 4.
...pK2: 1=1. ...t N) is detected. This peak extraction is performed using spectral normalized data lzi(10(K) K1
z −p K2 : l = 1 e ''・p N ) as follows.

zi, (8)-2i(6)〉0...
・(9)z1+2(K-21+,(8)〈0...αQ
When the above equations (9) and 01 are simultaneously satisfied, the spectral normalized data zi+, (K is the maximum value, (1+1
) indicates a local channel number, the local peak pattern Pi, and (6) are 1. ! :do.

The spectrum normalized data from i = 1 to (N-1) is sequentially checked to see if the α expression (9) is satisfied, and if it is, the local peak IJ? lX 7pi+
, (6) is set to "1", and if not satisfied, "0#" is set to 2.
It is converted into a value and stored in the peak pattern storage memory 403.

However, as a boundary condition, if i=l or "zi+1'Q-2i(PO(Q, then P, (
IQ = 1 ... α乃1 = (N-1) or Tsuzi +, (
'Q Zt([>Oo, P week: 1 ・(Ll), and if the condition is not satisfied, it is set to "0", it is binarized and stored in the local beakts mother tongue storage memory 403.

The local peak feature extraction unit 109 extracts the local peak pattern Pi (t=t
t2+...#N: に==に、... 2) A flowchart of extraction is shown in FIG.

In this way, the spectrum normalization in the data storage memory 401 is expressed as
e N ; ni== ni 1. ..., 2) is read every frame to obtain a binary local peak pattern. Standard Pattern Memo+J 110 superimposes multiple local peak patterns determined in advance to create one standard pattern for each category. This is where the created items are stored.

In order to normalize the difference in speech rate, the standard pattern calculates the average utterance length for each category, and linearly expands and contracts the time-direction frame length of the binary local peak pattern to match the average utterance length, from 1 to N channels. It is created as a superimposed, multivalued z4 tan. The standard/'P pattern memory 110 also stores a standard pattern length SL corresponding to each standard pattern.

The linear matching section 111 linearly expands and contracts the time frame lengths of the input local peak patterns inputted from the local peak feature extraction section 109 to the standard pattern length SL of each category in the standard four-tone memory 110. rear,
The initial similarity is calculated for each frame and the similarity is weighted by the average power. The initial similarity 5 (6) is calculated for each frame based on the following equation.

Here, P indicates the feature vector of the input local peak pattern in the first frame.

P=(pl(6), P2(6),...t, pNm)
...(1YD indicates the feature vector in the first frame of an arbitrary standard pattern.

D=(Dl(6)rD2■,...,DN(6))
...(6) Dt represents transposition, IIPII,
lID11 represents the norm of P and D, respectively.

Furthermore, when the average power of the voice is small, it is easily affected by noise, so the final similarity S is determined from the following equation 01.

Here, X/(K) is the average voice power determined by equation (8) above, max (X'(FQ) is the maximum value of the average voice power in the time-normalized section of the input voice, and SL is This is the standard pattern length.In other words, the final similarity S is the initial similarity S determined by the above equation 61 ([, the noise average power N, subtracted from the speech average power found by the above equation (8), or weighted by '(6))] This is the value.

The configuration of the linear matching section 111 is as shown in FIG.
701 is an input terminal for input local peak pattern, 702
is a standard pattern input terminal, 703 is an input terminal for audio average power, 705 is a standard 4-tang length input terminal, 706, 7
07, 708 are multiplication adders, 709, 710 are square root operators, 7 NO is a multiplier, 712 is a divider, 713 is a multiplication adder, 714 is a divider, 715 is a maximum value detector, 71
6 is a multiplier, and 217 is an output terminal.

The norm of the input local peak A tan P input from the input terminal 701 is multiplied by the adder 706 and the square root operator 709
Similarly, the norm of the standard pattern D is calculated by the multiplier adder 708,
It is determined by the square root calculator 710. Also, the multiplication adder 7
07, PD can be found as shown in the following equation.

Furthermore, the divider 712 determines the initial similarity of the equation (to) 5(8).

On the other hand, the audio average power X/(6) is input from the input terminal 703 and is determined by the multiplier/adder 713. In addition, the maximum value detector 715 calculates the maximum value of the time-normalized audio average power, which is multiplied by the standard pattern length SL in the multiplier 716, and this result is combined with the result of the α equation. is input to the divider 714 to obtain the final degree of similarity S, which is output to the output terminal 717 and sent to the determination unit 112.

In the determination unit 112, each standard no! The category name or category number of the standard pattern that gives the maximum final similarity for each pattern is outputted via the output terminal 113 and used as the recognition result.

(Effects of the Invention) As described in detail above, in the present invention, as a local peak extraction method, the least squares approximation straight line is used as the standard regardless of whether the sound is voiced or unvoiced, and the similarity is determined by the average voice power (=average electric power). Since weighting is performed using the Kerr noise average power), it is possible to compensate for the incompleteness of local peak extraction by the voice, especially in the low-level parts of the voice, even in voice mixed with noise.In other words, the peaks due to noise can be This has the effect of being easily recognized as a local peak.

[Brief explanation of the drawing]

FIG. 1 is a flowchart of the speech recognition method according to the present invention;
Fig. 2 is a flowchart of a conventional speech recognition method, Fig. 3 is a diagram for explaining the binarization of an input speech signal, and Fig. 4 shows the overall circuit configuration of an embodiment of a speech recognition device according to the present invention. 5 is a block diagram showing the circuit configuration of the frequency analysis section, FIG. 6 is a diagram for explaining voice section detection, and FIG. 7 is a block diagram showing the configuration of the local peak feature extraction section. Figure 8 is a flowchart for local peak extraction and tongue extraction, and Figure 9 is a diagram showing the circuit configuration of the linear matching section. 101...Input terminal, 102...Frequency analysis section, 1
03... Spectrum normalization section, 105... Average power memory, 106... Noise power calculation section, 107... Speech section detection section, 108-... Speech average power calculation section, 10
9...Local peak feature extraction unit, 110...Standard tactile memory, 111...Linear matching unit, 11
2... Judgment unit, 113... Output terminal. Patent Applicant: Oki Electric Industry Co., Ltd. Figure 1 Inventor's cap 1% Tanekyo power 1 phrase Flow child Dart Hatamai Ronso Figure 2 1 Figure 6
1. Indication of the case 1985 Patent Application No. 69053 2. Name of the invention Voice recognition method 3. Person making the amendment Relationship with the case Patent application Person's address (
105) 1-7-12-5 Toranomon, Minato-ku, Tokyo.
Target of amendment: “Detailed description of the invention” in the specification Format B
゛・,°”° ' 6. Contents of amendment (1) Specification @ page 10, line 1l, approximately rlomg.''
Correct it to "r 10 ms". (2) r KmK1. on page 13, line 1 of the same book. ...,
2; i=1. ..., 2)" is written as r(x=x1..., Km : i =11...
, N)". (3) r KECKL on the fourth line from the bottom of the same page in the same book,
Km 1J is corrected as [Kε(Kl #Km].

Claims (1)

[Claims] A process of frequency-analyzing input speech and extracting analysis data of a plurality of channels for each speech frame, and averaging all the analysis data of all speech frames in a predetermined noise interval to obtain a noise average power. a process of calculating the average power by averaging all the analysis data corresponding to the frame for each audio frame of the input audio; and calculating the average power and the noise average power for each audio frame of the input audio. A process of calculating the voice average power which is the difference between the spectral normalized and With respect to the analysis data, the presence or absence of a local maximum value along the frequency axis, that is, a local peak, is determined based on the least square approximation straight line, and if there is a local peak, it is set as "1".
A process of converting a binarized input audio into a local peak pattern with "0" in the case of no local peak, and a process of converting a plurality of standard patterns represented by a local peak pattern prepared in advance and the local The degree of similarity with the peak pattern is calculated for each normalized voice frame between the standard pattern and the input voice, and the degree of similarity is weighted by the voice average power, and the degree of similarity is calculated for each normalized voice frame between the standard pattern and the input voice. A speech recognition method comprising: a process of calculating a final similarity by accumulating the final similarity for each standard pattern; and a process of determining, as a recognition result, a category name of a standard pattern corresponding to the maximum value of the final similarity for each standard pattern.