JPS6073600A - Voice recognition equipment - 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 Technical Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device that is capable of efficiently and accurately extracting characteristic parameters of an input speech signal.

(b) Prior Art and Problems A speech recognition device includes a feature extraction section 2 which extracts the features of a speech signal inputted from a microphone 1 or the like, as shown in FIG. 1, for example. A voice pattern memory 5 in which the features obtained from the extractor 2 are registered as parameters. A recognition unit 4 is provided that recognizes input speech based on the feature parameters stored in the memory 5 and outputs the recognition result.

Here, as the configuration of the feature extracting section 2, there are a method of using a plurality of bandpass filters each having a different center frequency and wave number, and a method of using a correlator in place of these bandpass filters. However, the former method is somewhat disadvantageous compared to the latter method, as there is a concern that the circuit scale will increase or the amount of calculations will increase.

On the other hand, although the latter method is advantageous in terms of circuit scale or amount of calculation, there is a concern that the output characteristics may change if the correlation is taken for a long time. That is, the autocorrelation function is Y, the input audio waveform is 5 inches (2πft 〇), and the reference function is 5
In2ycfot and CO32πrot are defined by the following equation, assuming that the time used for waveform sampling is 8T.

The time (frame width) for taking the correlation is T=ΔTXN.

FIG. 3 shows the characteristics of equation (1) above. In other words, the voice waveform is converted using this characteristic to obtain the characteristic parameter. The longer the correlation time, the more accurate the transformed feature parameters will be. However, the time N for taking the correlation is 2N
, 3N .-mN), the output characteristic changes as shown in FIG. 4, which is a drawback.

(C) Object of the Invention The object of the present invention is to obtain accurate features in a method using a correlator in the feature extracting section so that the output characteristics do not change even if the time for correlating the input audio signal is changed. The object of the present invention is to provide a speech recognition device that can improve speech recognition performance by providing a correlation function that converts the parameters into parameters.

(d) Structure of the Invention In order to achieve the above object, the present invention repeats the correlation of the unit (corresponding to the time of △TXN) in the above-mentioned formula (1) multiple times, and calculates the result of cumulatively adding the respective correlation results. It is configured to perform feature extraction using .

The present invention will be explained in detail below using Examples.

(e) Embodiment of the Invention FIG. 2 is an L prism diagram showing the configuration of an embodiment of the present invention, and shows the configuration of the correlator in the extraction section 2 shown in FIG. In the figure, 21.22 is a multiplication section, 23.2
4 is a square calculation section, 25 is an addition section, and 26 is an addition accumulation section. Furthermore, Ys (△T) and Yc (△T) are the reference functions 5in2yrfQ△T and c shown in equation (1) above, respectively.
represents os2yr'fO△T.

That is, in this embodiment, the input audio signal IN is multiplied by the multiplier 21.2.
In step 2, take the product with the reference function at the sampling period △T,
Further, the square calculation units 23 and 24 perform a square calculation of each product. This calculates the correlation Y' of the unit frame. This unit of correlation output Y' is successively added and accumulated by the addition/accumulation section 26. Note that T is a delay circuit that buffers the cumulative result until the correlation output of the next frame is obtained.

In this way, the result of performing correlation for a unit frame (ΔTXN time) multiple times (for example, m times) is supplied to the feature parameter conversion unit (not shown) as the final output 0tJT of the number Y of correlators. . If this output 0tJT is shown in correspondence with the above equation (1), the following equation is expressed as (2πfΔT−1
10) Y2=cos2if, OΔT-i-sin (T to 2πf
-i 〇) Therefore, for example, the correlation output is repeated for the number of times (m+1)
By averaging the output characteristics as shown in FIG. 3, it is possible to obtain an output characteristic similar to that shown in FIG.

Furthermore, the period during which the correlation was taken is spread over a wide time series, allowing the characteristic parameters to be extracted with high accuracy.

(f) Effects of the Invention As detailed above (according to the present invention, the correlation time can be substantially lengthened without changing the output characteristics of the correlator in the feature extraction section), and the speech waveform can be further improved. This can be converted into accurate feature parameters, which in turn makes it possible to greatly improve the recognition rate.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of the configuration of a sound-m weaving device, FIG. 2 is a diagram showing an embodiment of the present invention, and FIGS. 3 and 4 are diagrams showing output characteristics of a correlator. Reference numeral 2 indicates a feature extraction section, 21 and 22 a multiplication section, and 26 an addition accumulation section. Also the 3rd
@, In FIG. 4, the horizontal axis represents frequency, and the vertical axis represents attenuation amount.

Claims (1)

[Claims] The apparatus includes a feature extraction section that extracts features of the audio signal waveform by making a correlation, the correlation section that takes a correlation between each of the temporally different signal waveform portions of the input audio signal and the reference signal; , a speech recognition device characterized in that the calculation results obtained from the correlation section are summed.