JPS61230199A - 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 peaks.

(Prior art) Conventionally, this type of speech recognition method was published in Journal of the Institute of Electronics and Communication Engineers, J68-A C1) (January 1985) p-78-
There was something described in 85. Figure 2 is a flowchart of a conventional speech recognition method using local peaks.The input voice is frequency-analyzed every 10m5ec by a group of 15-channel bandpass filters (see 1 in Figure 2), and the vocal cord sound source characteristics are analyzed. As a normalization method for individual differences, the voice spectrum is expressed logarithmically on both the amplitude and frequency axes, a least squares approximation straight line is obtained (see 2 in FIG. 2), and the difference is corrected. 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 Figure 3, for each frame (10m5ec), for each part where the amplitude is OdB or more, the channel that has the maximum value among those with an amplitude of 1/2 or more of each maximum value is identified as a local peak. Binarization is performed by setting the value to "1" and the other values to "0" (see 3 in FIG. 2). The number of channels of the band/base filter is 15, but when the slope of the least squares approximation straight line for the 16th channel is negative, it is considered to be a voiced sound and set to 1, and when the slope is positive, it is considered to be an unvoiced sound and set to "0". Then, add the sign of the slope (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 with a multi-value weighted average dictionary, the longer pattern is linearly stretched and matched in the time direction, calculations are performed based on a certain degree of similarity, and a standard pattern that gives the maximum degree of similarity is selected. The category name is used as the recognition result (see 6 in Figure 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 is a problem in that noise peaks are likely to be picked up as local peaks in silent parts, leading to an increase in erroneous recognition.

The present invention eliminates the problem of the prior art in that in a noisy environment, noise peaks are extracted as local parts of speech, especially in silent parts, resulting in an increase in false recognition, and a speech recognition method that is highly resistant to noise is provided. (Means for solving the problem) The speech recognition method according to the present invention first performs frequency analysis of an input speech signal into analysis data of multiple channels for each speech frame. Next, the average value of the analysis data, that is, the average power, is calculated for each audio frame.

In addition, each analysis data is spectral normalized using the least squares approximation straight line of the audio spectrum, and the local maximum value along the frequency axis, that is, the local The presence or absence of a peak is determined and converted into a binary local peak pattern in which the presence of a local peak is set to "1" and the absence of a local peak is set to "0". Then, by calculating the degree of similarity for each voice frame between the plurality of standard/IP tongues and the local peak pattern of the input voice, weighting this degree of similarity by the above-mentioned average power, and accumulating it over all voice frames. Perform the final similarity calculation. Then, the category name of the standard pattern corresponding to the final similarity that gives the maximum value is determined to be the recognition result.

(Function) The present invention calculates the final similarity by weighting the similarity between each standard pattern and the local peak pattern of the input speech by the average value of the frequency analysis data of the input speech, that is, the average power. (Embodiment) Figure 1 is a flowchart of the speech recognition method according to the present invention. The signal is frequency analyzed by a frequency analysis Stellar flZ every l Q m5ec (frame period), and logarithmic values as analysis data of multiple channels are output for each audio frame. Next, average power calculation step 1
2, the average power is calculated by averaging all logarithmic values within each audio frame. Further, the logarithmic value of this frequency analysis data is spectral normalized using the least square approximation straight line of the audio spectrum in the audio frame to which this logarithmic value belongs (Stella f13). Next, the presence or absence of a local maximum value along the frequency axis, that is, a local peak, is determined for the spectrum-normalized logarithm value using the least square approximation line as a reference, and if there is a local peak, it is set as "1".
and no local peak.

Converts to a binary input local peak value with "0" in the case of .

The standard pattern 15 is created in advance by linearly expanding and contracting a plurality of binary local patterns to the average utterance length for each category and adding them.

In step f16, the binary input local peak pattern, each standard 1? linearly expands and contracts to the tongue frame length;
The degree of similarity for each voice frame is calculated, the degree of similarity is weighted by the average power output from the average power calculating Stella f12, and the weighted degree of similarity is accumulated over all voice frames to obtain the final degree of similarity. Then, the category name of the standard pattern that gives the maximum final similarity is output as the recognition result.

Next, one embodiment of a speech recognition device using the speech recognition method of the present invention will be described.

FIG. 4 is a block diagram showing the overall circuit configuration of one embodiment of the speech recognition device according to the present invention.

In FIG. 4, 101 is an audio input terminal, 102 is a frequency analysis section, 103 is a spectrum normalization section, 104 is an average power memo'), 105 is a voice section detection section, 106 is a local peak feature extraction section, and 107 is a standard/ fmethanemo'),
10B is a linear matching section, 109 is a determination section, and 110 is an output terminal for the recognition result. The operation will be explained below. First, an input audio signal converted into an audio signal by a microphone or the like is input to the audio input terminal 10, and is passed through the frequency analysis section 10217 (thereby extracting a frequency spectrum. Frequency analysis section 1020 circuit) The configuration is shown in FIG.

In FIG. 5, 201 is a preamplifier (Amp), ! 02
1tt/4-pass filter group (BPF-1, . . .

BPF-N), 203 is full wave rectifier details (DET-1,
....

DET-N), 204 is a low-pass filter group (I, P
F-1°..., LPF-N), zos is an analog multiplexer, 206 is an AD converter, and 207 is a ROM for logarithmic conversion. The band A filter group is arranged from 1 to N with center frequencies at equal logarithmic intervals, starting from the lowest N frequency, hereinafter the first channel, ..., the Nth channel.
Written as a channel. In this embodiment, N=22.

The output of the loaf 9 filter 8204 is switched by the multiplexer 205, converted into a digital quantity by the AD converter 206 at a sampling period of about 10 nu+, converted into a logarithmic value by the logarithmic conversion ROM 207, and then sent to the spectrum normalization unit 103. will be sent to.

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 nine banks of digital filters are used.

The logarithmic conversion ROM 207 outputs audio spectrum data for N channels of the second frame as X, (6
), ..., X. The spectrum normalization unit first calculates the average value X(6) for each frame using the following equation (1).

The following discussion will be referred to as the "average power" in the frame. The average power is stored in the average power memo I7104, and then sent to the voice section detection unit 105.

Speech interval detection can be performed, for example, by using two thresholds E for the average power.
, , K2(E,)K2). In other words, as shown in Figure 6, the average power exceeds K2, and then E
When E is exceeded without becoming smaller than 2, let the point beyond K2 be the starting point. In the same way, the time axis is reversed and 2 is obtained at the end, and (K1 s K2 ) is set as the voice interval. *1=L
・-, N) is normalized using the least squares approximation straight line according to the following (2) 2. The difference between the spectral data X1 (6) and the least squares approximation straight line is calculated using the following equation (3) to obtain spectral normalized data Zi (6).

Zi(X)=Xi(IQ (A([・i +B(K)
) (K=to 4...t K2; 1==l p...
・, N)...(3) A(K), B in the least squares approximation straight line of the above equation (2)
(6) is given by the following equations (4) # (5).

The calculation of equation (3) is performed when A(6) is positive, that is, in most cases, even for unvoiced sounds. In the time direction, the above-mentioned equation (3) is calculated for all frames K before the voice section is detected.

Next, the local peak feature extraction unit 106 extracts Zi(6) (K==4,...,2) from the start to the end of the voice.
;i=1. . . . p N ) are sequentially processed frame by frame.

The detailed configuration of the local peak feature extraction section 106 is shown in FIG.

In FIG. 7, 401 is a data storage memory, 402 is a peak extraction section, and 403 is a local peak extraction storage memory.

The data storage memory 401 contains spectral normalized data Zi(6) (K==) from 1 at the beginning to 2 at the end.

...g K2; i = 1 a 2 m...,
Zt([0(K
= K1 t ”' a K2 ; 1=1...
, N) is detected. This peak extraction is performed using the spectrum normalized data ZiaQ (i =
l-N) as follows.

zi+ , "-2i('0 >O・" (6) zi+
2(6)-2i, (6)〈0...(
7) (6) 1 When the equation (7) is simultaneously satisfied, the spectral normalized data ziya, (6) is the maximum value,
(t+x) indicates a channel number with a local peak, and local peak/mother tongue P1+1(8) is assumed to be 1.

The spectrum normalized data from i = l to (N-1) is sequentially checked to see if it satisfies the above formulas (6) # (7), and if it is satisfied, the local peak pattern Pi,
(6) is set as ``1'', and if it is not satisfied, it is set as ``0'' and 2
It is converted into a value and stored in the peak pattern storage memory 403.

However, as a boundary condition, i = 1 or' zt+1GOZ 1(K) < O(
When D, P, (K) = 1 - (8)i = (N-1)
Katsuz,, (When [-z, GO'> 0, P-=1...
(9) If the condition is not met, it is set to "0", binarized and stored in the local beak tongue storage memory 403.

In the local peak feature extraction unit 106, the local peaks 4 tan Pi(8) in the certain frame are
(+ = 1 p 2 e..., N) A flowchart of extraction is shown in FIG.

Spectral normalization 7'-I in data storage memory 40
Zi(K)(i=1 t 29-9N'*to=to1t-
tK2) is read every frame to obtain a binary local peak gloss.

The standard pattern memory 107 stores a plurality of pre-determined local peak patterns superimposed to create one standard pattern for each category. In order to normalize the difference in speaking speed, the standard Tsukutan calculates the average speaking length for each category and asks for the binary local peak iJ? The frame length in the time direction of the tongue is made equal to one average utterance length, linear expansion and contraction are performed, and channels 1 to N are superimposed to create a multivalued pattern. In addition, this standard pattern memory 1
07 has the standard pattern length S corresponding to each standard/yatan.
L is also stored.

In the linear matching section 108, the time frame length of the input local peak patterns that are input multiple times from the local peak feature extraction section 106 is set to standard/J? After aligning the standard pattern length SL of each category in the pattern memory 107 by linear expansion and contraction,
The initial similarity is calculated for each frame and the similarity is weighted by the average power. The initial similarity 5 (8) is calculated for each frame based on the following formula 01.

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

P = (P, (fQ, P2(IQ, -, P, H
) (1O indicates the feature vector in the first frame of an arbitrary standard pattern.

D=(D, (6)t D2GOt..., DNoo)
...to) ()1 represents transposition, IIPII
, IIDI+ represent the norms 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 (2).

SL-max(X(K)) Here, X(6) is the average power found by equation (1), max(
X(fQ) is the maximum value of the average power in the time-normalized section of the input audio, and SL is the standard pattern length. That is, the final similarity S is a value obtained by weighting the initial similarity S@) determined by the α1 formula by the average power.

The configuration of the linear matching section 108 is as shown in FIG.
701 is an input terminal for input local peak pattern, 702
is the standard pattern input terminal, 703 is the average power input terminal, and 705 is the standard iJ? Tongue length input terminal, 706,7
07, 708 is a multiplication adder, 709.710 is a square root operator, 711 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 717 is an output terminal.

The norm of the input local peak pattern P from the input terminal 201 is calculated by the multiplier/adder 706 and the square root calculator 709 as shown in the following equation.

Similarly, the norm of standard pattern D is multiplied by adder 708,
It is determined by the square root calculator 710. Also, the multiplication adder 7
07, PDt can be found as shown in equation (b).

Further, the divider 712 divides the initial similarity 5 of the formula 01 into
(Kl is determined. On the other hand, the average power XE is input from the input terminal 703, and the multiplier/adder 713 determines the maximum value of the audio average power. ' A final similarity S is determined by a divider 714 using this result and the result of the oQ formula, and is output to a terminal 717 and sent to the determination unit 109 .

The determining unit 109 outputs the category name or category number of the standard pattern that gives the maximum similarity for each standard pattern via the output terminal 110.

(Effects of the Invention) As described in detail above, in the present invention, as a local peak extraction method, a least squares approximation straight line is used as a reference regardless of voiced or unvoiced sounds, and similarity is weighted by average power. Therefore, it is possible to compensate for imperfections in local peak extraction due to voice, especially in low-level parts of the voice, even in voice mixed with noise.In other words, peaks due to noise are less likely to be regarded as local peaks due to voice, which is good. has a discernible effect.

[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 is the overall circuit configuration of one embodiment of the 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. 8 is a flowchart of local peak A'tan extraction, and FIG. 9 is a block diagram showing the circuit configuration of a linear matching section. 10 No. Input terminal, 102 Frequency analysis section, 1
03... Spectrum normalization section, 104... Average power memory, 105... Voice section detection section, 106... Local peak feature extraction section, 107... Standard iJ? Tan memo'), lag...linear matching section, 1o9...
- Judgment unit, 110...output terminal. Patent Applicant: Oki Electric Industry Co., Ltd. Fig. 1 Blind P recognition method of the present invention Fishing 70-Natto Cotton Nail Fig. 2 RL Blade' Immediately on the 6th Fig. 5 1!'1 Purchase of the number analysis department 102 Fig. 6 Responsibility Tomoe - Jiebo / IM@1 Association No. 71 0 - Calbee 7 Special Walk Work n + 064 Precision procedure correction (Spontaneous) 1. Indication of the case 1985 Patent Application No. 070035 2, Name of the invention Speech recognition method 3, Person making the amendment Relationship with the case Patent application person office (
Address: 105) 1-7-12 Toranomon, Minato-ku, Tokyo Name (029) 'x Chuo Denki Kogyo 1-piece company representative
Sukeryu Naga Hashimoto Nankai 4, Agent address (〒105) 1-7-1, Toranomon, Minato-ku, Tokyo
2 No. 5, Subject of amendment: "Detailed Description of the Invention" column in the description and drawing "Figure 8" In the fifth line, “sampling period 10
The phrase ``with a sampling period of about 10 m5'' has been corrected to ``with a sampling period of 10 m5''. (2) In the corner of page 17 (16) of the circular. (3) In the second page of page 17 (17) of the circular. (4) The drawing "Figure 8" is corrected as shown in the attached sheet. that's all

Claims (1)

[Claims] 1. A process of frequency-analyzing input audio and extracting analysis data of multiple channels for each audio frame, and averaging all the analysis data corresponding to the frame for each audio frame to obtain an average power. a process of calculating the spectrum of the analysis data using a least squares approximation straight line of the audio spectrum in the audio frame to which the data belongs; and a process of performing the least squares approximation on the spectrum-normalized analysis data The local value of the binarized input audio is determined by determining the presence or absence of a maximum value along the frequency axis, that is, a local peak, using a straight line as a reference, and setting it as "1" if there is a local peak and "0" if there is no local peak. A process of converting into a peak pattern, and a process of converting the similarity between a plurality of standard patterns represented by local peak patterns prepared in advance and the local peak pattern of the input audio into a normalized audio frame between the standard pattern and the input audio. a process of calculating a final similarity by weighting the similarity by the average power and accumulating over all the normalized audio frames; and a process of calculating the final similarity for each standard pattern. A speech recognition method comprising: determining a category name of a standard pattern corresponding to a maximum value of degree as a recognition result.