JPS592100A - Voice recognition system - 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 sequentially calculates the degree of similarity between the characteristic pattern of the voice to be recognized and the characteristic pattern of a plurality of types of voice alterations determined in advance by J1.
The present invention relates to improvements in speech recognition methods that perform recognition judgments based on calculations, and particularly relates to speech recognition methods characterized by pattern mapping methods.

Conventionally, recognition is determined by sequentially calculating the degree of similarity between the characteristic pattern of the voice to be recognized and the characteristic patterns of multiple types of rI voices registered in advance. The pattern matching method using the dynamic lateral stroke method in the 1-voice 3.3 recognition device is as follows:
It is known as a powerful means for changing the time axis of vocalization [j, 1211].

However, in the conventional pattern matching method, it is necessary to process the input voice and the registered fH voice into 71 frames according to the length of each frame.
It required a lot of time, both in processing time and in the standard pattern method.

The present invention has been made in view of the above two points, and aims to provide a speech recognition method capable of significantly reducing both processing time and memory occupancy for standard patterns. I'm a big fan.

In order to achieve this object, the present invention provides a human-powered speech recognition device that performs recognition judgment by successively calculating the degree of similarity between a speech feature pattern to be recognized and a plurality of pre-registered speech feature patterns. audio,! ; Both with the registered voice and the analysis frame of Ml at a certain time,
Extract and linearly expand and contract the number of flakes J according to the length of occurrence time to obtain the constant flake l, number of °C111, and then dynamically, f +
It is configured to perform recognition determination by pattern matching using a drawing method.

That is, in the present invention, both the human-powered 7h and the registered high-pitched voice are linearly expanded to equal length a? It is characterized by performing the mano-chinking process JJli by dynamic programming without performing any processing such as returning to the original number of planes. The memory requirement for storing is significantly reduced.

Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings.
I'll do it.

FIG. 1 is a schematic block diagram showing an example of a speech recognition device embodying the present invention, and FIG. 2 is a recognition operation flow diagram of the 13-voice recognition system of the present invention.

In Figure 1, 1 is a microphone that picks up voice input, 2 is an amplifier, 3 is an A/D converter that converts human input voice into a digital signal, and 4 is a person Q. Extract! Li￥ signature extraction circuit, 5 is the human-powered j′↑ voice? A voice section detection circuit that detects a voice section, 6 is a microcosm sensor (CI) u ) that performs various calculations 1'tW, 7 is an input pattern memory, 8 is a pattern memory, 9 is a dynamic A matching (DI' matching) circuit based on the old painting method, and 10 an output circuit for outputting 3.3 recognition results and the like.

As one method of speech recognition, there is a method in which a feature pattern of a voice to be recognized is registered in advance, the feature pattern is matched with a feature pattern of input speech, and the most similar pattern is output as a recognition result. In this method, generally speaking, speech is analyzed with a certain analysis frame length7, the feature vector of that frame is extracted for the length of the speech, and the time series of the feature vectors is used as the feature pattern of the speech.

For example, in Figure 3, the voice waveform a for the utterance of "Mishima"
, the analysis frame width l], the frame number C1, the feature vector of each frame)/in the figure, the autocorrelation coefficient).

In Fig. 1, the sound μm waveform a microphone o is uttered.
+h nl f jfi, AGC i, 1 amplification a:'
j 2-c 6 dt<, z, C.

The high range is emphasized and amplified. 1° which is an analog signal
Conversion from 1 voice signal 7j to digital number is done in A/D conversion product 3, +j (+always, with sound ijY recognition, sunblink 2 with 8 to 241NIz (e.g. 8KIIz) , each sunblink f (i'i' is 6 to 1
2 Represented by a Hiso I- (e.g. 8 bits) digital signal.

On the other hand, the length of the human-generated voice is detected by the voice section detection circuit 5. The sound during this audio interval is 1
is analyzed frame 7ii by the feature extraction circuit.

The feature vector 1f″L1,
In this way, a feature vector time series of multiple frames is formed for human speech (step 112 in Figure 2).
). For example, when using an autocorrelation coefficient as a feature vector, as shown above (a signal quantized to 8 bits with 8 KHz numbering is quantized to 8 bits with 128 data pieces as one frame and 24 bits of data up to the 15th order). autocorrelation function R
(m) is calculated, and then, for example, the 1st to 8th orders are divided by the 0th order value to obtain an autocorrelation coefficient with 4-bit accuracy, which is used as the 'i, I, ; signature parameter.

The characteristic pattern of the plurality of flights is converted into a number of fixed flights of, for example, 16 rooms by linear expansion and contraction (SEDEP! 13), and is stored in the standard pattern memory 8 at the time of registration, and stored in the input JJ Bakun memory 7 at the time of recognition.

The divine method can be considered as the linear expansion/contraction method, but in the embodiment of the present invention, an example of expansion/contraction to 16 frames is shown below.

Assume now that there are n frames of sections that are determined to be speech by the speech section detection circuit 5. Also, the feature vector in each frame is expressed as 1 (1), and the expanded and contracted feature vector)/
If V is represented by To(i), linear expansion/contraction processing is executed by the following process.

(b) The first frame is unconditionally adopted.

h (1) = a (1) (mouth), the remaining (n-1) frames are divided into 16 sections.

([]dcousdself) That is, by processing (a) and (b) above, for 1)il 15 of 1G intervals, σfj frame J,
is set as each feature vector l-71/, and one frame can be expanded and contracted into 11 frames and 16 frames.

For example, j゛ (in the example in Figure 3, an O mark is added to the frame number υ) indicates the frame that will be adopted when the frame is expanded or contracted to 16 frames and 7 is reached.

Figure 4 (a) and (1)) each have one voice section.
L20 Freno 1.13 Frame R is the same as linear expansion and contraction 1 at frame J. Fl is the original frame number J-, p2 is the frame number at the time of expansion and contraction. . ! ,
Freights marked with an "X" in the title number are those that will not be adopted.If the number of frames is less than 6 frames, the same frame J may be adopted twice.

c (7) After the feature pattern of the 1) frame extracted by the above process is linearly expanded/contracted to 16 frames, the characteristic pattern is expanded/contracted to 016 frames, and then the characteristic pattern is converted to 11 by the P matching circuit 9. , dynamic IN image θ, recognition by matching, so-called DP matching, is performed (step n4), and the result is output to the output device 10.
(Nutenobu n5).

There are many adaptation formulas for DP matching, but for example, the following formula using the path shown in FIG. 5 can be used. Of course, it is possible to perform 74 JQ using other formulas, and the present invention is not limited to the following formula.

(a), g(1,l)~26(1,1)Ii
−j l≦r, i=2 to 1; d(i, j) is the feature vector a (i) of the first frame of human speech;
'H?' of one j-th frame of the 1m quasi-pattern?
? a represents the distance of vector tl+(i).

r represents the size of the matching window, which is 16 frets.) If the linear offset i is to the extent of 0, it can be made as small as r-1 to 2, so Iil'-1": 4
iX can be reduced. FIG. 6 shows an example of the first flash pass.

1 horn, 7": The recognition judgment based on the old "matching" is executed by the I) P manchiking circuit 9, and the recognition result is output from the output device 1.0.

As described above, according to the present invention, ge, human sound 7rr,:
With ``'', both X WA t:f mountains are analyzed for a certain time, feature fit is extracted, and all extraction (~, the number of frames is linearly expanded or contracted according to the uII length at the time of utterance, and the fixed number of frames and L7,
Since it is a l'71 configuration in which recognition-specific car 1 is then performed by pattern matching using the dynamic 1-11 drawing method,
In addition to being able to significantly reduce recognition processing time by one bit compared to recognition using conventional pattern matching,
The ability to significantly reduce the amount of memory required to store feature patterns is i-+J.

[Brief explanation of the drawing]

FIG. 1 shows the configuration of a speech recognition device implementing the present invention. Figure 2 is a recognition operation flowchart of the recognition method of the present invention, Figure 3 is an example of an ei speech waveform and its analysis waveform, ]':, Figure 4 is a linear diagram of the analysis frame J. Figure 5 is a diagram for explaining expansion and contraction; Figure 5 is a diagram showing an example of a path in matching (January); Figure 6 is a diagram showing an example of an optimal path by matching (old).

Claims (1)

[Claims] 13. Recognition determination is made by sequentially performing calculations on the degree of similarity between the characteristic pattern of r°1 voice to be recognized and the characteristic patterns of multiple types of voices registered in advance 7:' In the I voice recognition device, features are extracted from both the human voice and the registered voice using Freno analysis for a certain period of time, the number of frames is linearly expanded or contracted according to the utterance time 1φ, and then A speech recognition method characterized in that recognition determination is performed by dynamic programming/< turn matching.