JPS62283400A - 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 3. Detailed Description of the Invention (Field of Industrial Application) The present invention relates to a speech recognition method, and particularly to a speech recognition method that performs word speech recognition using local peaks.

(Prior art) Conventionally, as this type of speech recognition method, the Acoustical Society of Japan Proceedings l-4-12 (March 1985) P, 23-24
There were three articles published in .

FIG. 3 is a flowchart for explaining a conventional speech recognition method using local peaks. In the conventional method, the input voice is converted into 12 H2, 12-bit A/D conversion in step 30 (hereinafter the step is denoted by S), and then N channels (where l is the number of channels, i=1, . . . ) N) For example, a 22-channel digital bandpass filter (B
PF) Bank performs frequency analysis every 10 msec and converts the input pattern to j3 (S31). Next, local maximum values in the frequency and axial directions, that is, local peak features, are extracted from the obtained input pattern (S32).

Amplitude and frequency are used as local peak feature extraction criteria.
Find the minimum knee approximation line expressed logarithmically. The channel corresponding to the peak that exceeds the obtained least squares approximation straight line is set to "1" as a local peak, and the remaining channels are set to "0" to perform binarization.

#11 Whether the input voice is for registration or recognition target
According to (33, 3), if the pattern is for a composite line, a standard pattern is created by adding the binarized patterns described above (May).

On the other hand, if the pattern is for recognition (hereinafter referred to as input pattern), the similarity with the standard pattern is calculated and the category name of the standard pattern that gives the maximum similarity is output as the recognition result (S35).

The degree of similarity is defined by equation (1) when the second frame in the large pattern for recognition and the m-th frame in the standard pattern are made to temporally correspond.

Note that 2 and m are arbitrary positive integers.

Sl,=Xt 2! / (IX'l llz, i
) ... (1) Kokote'lL = (X+A, X
Z! , ",
l Xtfi,ilZ,,l is 6xQ,,2. ,
, Nonormuwo was an expression. That is, .

(Problem to be solved by invention) However, in the conventional method described above, (1) and (
2) As can be understood from the formula, multiplication, square root operation, and division are performed in similarity calculation, so the operation speed is slow.
There was a problem in that the circuit configuration for arithmetic processing was complicated and large.

The object of the present invention is to eliminate the above-mentioned conventional problems and to provide a method for recognizing running speed that uses cumbersome calculations and is superior in operating speed.

(Means for Solving the Problems) In order to achieve the first objective, the speech recognition method according to the present invention uses a method for calculating similarity in speech recognition using local peaks. Whether it is a standard pattern or a multi-valued value, the standard pattern is found by the sum of the diameter lines when the input pattern takes the value "1", so the value of the standard pattern is a small value from 0 to the number of q records. Based on this fact, the similarity calculation is performed without performing a root operation or division.

Therefore, according to the present invention, in the local peak feature extraction process, only a predetermined number of low frequency local peaks are extracted as local peaks. In this case, the low frequency band means the lowest frequency channel number, that is, the first channel when numbering channels in frequency analysis from the low frequency side. Further, in this case, the predetermined number refers to a number that does not cause a decrease in the recognition rate in recognition processing, and is preferably seven, for example.

Next, according to the present invention, the following three counting processes are performed in the similarity calculation process. In the first counting process, the number in which the element X+ (i=1.2, . . . , N) of the input local peak vector of the recognition target pattern, that is, the input pattern is “1” is counted.

In the second counting process, the element X of the input local peak hector
, is "1 pa", the sum of the element Z1 of the corresponding standard pattern vector is calculated.In addition, in the third counting process, the sum of the element Z1 of the standard pattern vector or the element Z when it is other than "0" is calculated.
, calculate the sum of squares.

Furthermore, in this similarity calculation process, by referring to the table using the nine values obtained in these three counting processes as an address, the similarity value is read from this duple and output. In this table, similarity values corresponding to addresses are stored in advance and formed as a table. Further, the above-mentioned standard pattern is also stored in the memory in advance.

(Operation) As described above, according to the present invention, the input local peak vector of a large Kapatan is set as a binary value, and the standard pattern hector is set as a multi-valued value, and furthermore, the magnitude of the value of this standard pattern vector is from 0 to 0. The maximum value is the total number of times the input local peak vector takes "1", and the number of local peaks extracted from the input pattern is set from low to L!
The number of calculations for calculating the similarity can be significantly reduced by setting the number to a small value that does not cause a decrease in the recognition rate.

Further, according to the present invention, there is a method of reading the corresponding similarity value from the table using the value obtained by multiplication and addition of the multiplication results without performing division or eastern root operation.

For this reason, the calculation speed processing for calculating the similarity can be simplified, and therefore the calculation speed can be increased. Furthermore, based on the simplification of arithmetic processing, it is possible to simplify and downsize the device for implementing this traveling speed recognition method.

(Embodiments) Hereinafter, embodiments of the present invention will be described with reference to the drawings.

In the following embodiments, a case will be explained in which the number of channels is 22 instead of 1, but it should be understood that the present invention is not limited to this.

FIG. 1 is a block diagram of the overall configuration of an embodiment of the present invention, in which 10 is an audio input terminal, 12 is an A/D converter, and 14 is an audio input terminal.
is a band pass filter (BPF), +6 is a local peak extractor, 18 is a voice section detector, 20 is a memory, 22
is standard pattern memory, 24 is Lf/11 section detection section 1
This is a control unit for controlling the memory 20 and the standard pattern memory 22 using information signals from the standard pattern memory 8. 26 is a similarity calculation unit, which includes a counting unit 28, 30, 32, and a table R.
Similarity daple 3 formed in AM or table ROM
4 and a matching section 36. Furthermore, 38 is a determination section, and 40 is an output terminal. In terms of the heart configuration, preferably 16, 18, 28, 32 and 36 in FIG.

Next, the speech recognition method of the present invention will be explained with reference to the apparatus block diagram in FIG. 1 and the flowchart of the spoken word A in FIG. 2.

First, audio is input to the audio input terminal IO (520), and is converted into digital audio by the A/D converter 12 (S21), as in the past. This audio is N depending on the BPF.
Channels are analyzed into 22 channel frequencies in this example (S22). The output after the frequency analysis is detected as a voice section by the voice section detection section 18 (S23), and then sent to the local peak extraction section 16, where a peak exceeding the least squares approximation straight line is set as "1" as a local peak. The remaining values are set to “0” and binarized,
Local peak vector X= (X,,X2,...,
×22) is extracted (S24). Note that local peak extraction is performed within the voice section determined by the voice section detecting section 16 and stored in the memory 20. Here, the start end information of the voice section is assumed to be I, and the end point +11 information is assumed to be IE.

The voice section detection unit 18 described above detects the BP of a certain specific band.
A voice section is determined by determining voice and non-voice frames based on the value of the F output. The voice section determination method is based on the fact that signal energy increases when voice is input, and is performed by comparing a certain threshold value with a BPF output value. There are also a number of other methods, including a method that supplementarily uses other information (for example, the number of zero crossings), but their explanation will be omitted since they are not the purpose of this invention. Furthermore, in this embodiment, the information (5) and (8) of the voice section detection unit 18 is sent to the control unit 2.
4, and the control unit 24 sends the input second frame (l
(values within the voice interval) and the m-th frame of the standard pattern (described in detail later) are transmitted to the memory 20 and the standard pattern memory 22, respectively.

Next, calculate the similarity! A similarity equivalent to that shown in equation (1) described in 1526 is calculated (S25). This is achieved by the method shown below. The calculation unit 28 calculates the mth frame (ffi=[I19,
I, ]) of the input local peak vector Xt
+-(i=1.2, . . . , 22) or “l” is counted (first counting process). Since the input local peak extractor is binary and I×1=1, this is: l xbi2=XL ty =ΣX822...
・(3)mr will be multiplied by Δ1.

The binary input local peak vector
(7) Elements x, A (i=t, 2,..., (2) are "1"
”, the element z of the standard pattern hector, , (i=
+i, 2, ..., 22)'6 (second counting process). That is, since X and l are "1" or "0", they are counted.

The counting unit 32 calculates element Z1. of the standard pattern vector 2m.
．． The sum of squares of the elements Z +++ other than "0" is calculated (third counting process). That is, 12.12=7,2°; Σ21ffi...
(5) 1; A is counted.

In this embodiment, the local peak extractor 16 extracts a predetermined number of local peaks from the low range, for example, up to seven local peaks from the 22 channels. This may be because the recognition rate does not decrease at all even if the number is limited to seven, and the memory reduction effect (described later) is large, or the number can be set to an appropriate number that does not cause a decrease in the recognition rate.

In this example, the drum value of the number of local peaks is set to 7, so equation (3) has a value of 1 to 7, which is expressed in 3 bits, and equation (4) has a value of 0 to 2+(=3X7), which is 5. Expression (5) has a value of 1 to 63 (=3'×7) and can be expressed in 6 bits.

Therefore, formulas (3) to (5) are combined into 14 bits (=3+
5+6).

The similarity value 5L-o stored in the table 34 is read out using the 14-bit value represented by the values obtained through these three counting processes as an address (S26), and this table 34 is written as (3). ～ku5)
This is a conversion table for calculating the value of equation (1) from the equation, and in this embodiment, Pth 8-bit similarity values are stored. The capacity of the table 34 is a 14-bit address, i.e.
It only takes 16K bytes. The similarity value in equation (1) was set to 8 bits because the performance did not deteriorate even with 8 bits.

Based on the similarity value Sp- obtained as above, the matching section 36 calculates the overall similarity S (S27).

In this case, there is a method of non-linear correspondence between 1 and m, or in this embodiment, it is assumed that they are made to correspond linearly. At this time, the overall similarity S is determined by the following equation (6).

Here, k is a certain standard pattern number (k=1.2,...
, K) and Lk are the lengths of the second standard patterns stored in the standard pattern memory 22, and the total number of standard patterns is shown, respectively.

As described above, for all the standard patterns Sk (
k=1.2,...), and '! -+1 constant part 3
Send to 8. In the determination unit 38, the maximum similarity is 0, that is, ko=arg max S+.

)≦kiK. Select and send to output terminal 1'-40 (528)
.

The invention is not limited to the embodiments described above. For example, in the above-described embodiment, an example of 22 channels has been described, but the present invention can be applied to any number of channels. Furthermore, since the present invention is a voice recognition method, the device configuration for carrying out the method is not limited to the above-described embodiment.

(Effects of the Invention) As is clear from the above description, according to the present invention, by utilizing the properties of local peak patterns of speech, similarity calculation can be realized with a small memory capacity and a simple counting section. This eliminates the need for square root calculations and divisions, which were conventionally required, and it becomes possible to perform similarity calculations at high speed with a simpler configuration than conventional methods. A device implementing this invention can be realized as a small-sized speech recognition device.

[Brief explanation of the drawing]

Figure 1 is the audio of this invention! 1 is a block diagram showing an example of the configuration of a speech recognition device for carrying out the present invention to provide an explanation of the method; FIG. 2 is a flowchart to explain the speech recognition method of the present invention; and FIG. It is a flowchart used to explain a recognition method. 10... Audio input terminal f, 12... A/D conversion number 14... Band pass filter bank 16... Local peak extractor 18... Voice section detector, 20... Memory 22...
...Standard pattern memory 24...Control unit, 26...Similarity counting unit 28-32...Counting unit, 34...Table 3
6... Matching section, 38... Judgment section 40...
・Output terminal. Patent applicant Oki Electric Industry Co., Ltd.@? 42-tnI? SQ '4't * La'4
V') 3L h,"@Fig.3

Claims (1)

[Claims] (1) A process of frequency-analyzing the input audio to obtain an input pattern, finding a least squares approximation straight line from the input pattern, and setting the peak exceeding this to "1" as a local peak, and setting the others to "0". A local peak feature extraction process that performs binarization by setting the above-mentioned In a speech recognition method including a process of determining input speech according to similarity, the local peak feature extraction process has a function of extracting only a predetermined number of local peaks from the low range, and the similarity calculation process has a function of extracting only a predetermined number of local peaks from the low range. a first counting process that counts the number of "1" elements in the local peak vector; a second counting process that calculates the sum of the elements of the corresponding standard pattern when the input local peak vector is "1"; The third counting process calculates the sum of squares of the elements when the elements of the pattern vector are other than "0", and the values obtained from these first, second, and third counting processes are used as addresses, and similar values are calculated from the conversion table. A speech recognition method characterized by including processing of reading and outputting a degree value.