JPS59231599A - 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 for recognizing speech based on the results of spectrum analysis of speech.

Conventional configuration and problems Generally speaking, to recognize speech, it is necessary to perform spectrum analysis or equivalent processing every fixed period of time (10 to 20 m5ec) from the speech electrical signal waveform converted by a microphone, etc. This requires analysis and a discrimination process that uses the analysis results to discriminate between categories. However, when performing speech recognition under noisy conditions, recognition performance deteriorates. This is because the audio signal is contaminated by noise, and the matching distance from the standard audio is longer.

In conventional speech recognition methods, in order to remove the contamination of the speech signal due to noise, the analysis results are passed through a low-pass filter or smoothed by time averaging. Such processing removes local spike-like noise that appears in the analysis results when noise is added to the audio signal.

However, the smoothing according to the conventional example described above not only removes noise components, but also deforms the waveform of the original analysis result, making it difficult to match with standard speech in discrimination processing, and reducing recognition performance in noisy environments. There was a problem that the performance did not improve much.

Purpose of the Invention ``The present invention eliminates the problems of the above-mentioned conventional example, and removes the superimposed noise component without preserving the original waveform of the analysis result, thereby preventing the deterioration of recognition performance due to noise. The purpose is to provide a speech recognition method that prevents

Structure of the Invention In order to achieve the above object, the present invention removes noise components superimposed on the audio signal by smoothing the audio analysis results by median averaging, and uses this smoothed output for category discrimination. This information is used to obtain the effect of preventing a decline in recognition under noisy conditions.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

In FIG. 1, an audio signal converted into an electrical signal by a microphone 9 is input to a filter bank 2 via an amplifier circuit 1. The filter bank 2 includes a plurality of bandpass filters 2.
01,202. ...207L and the rectifier circuits connected to each 211.212+...21? L, integrating circuits 221, 222. ...22n. The outputs 23□, 23□, . The smoothed signal 6 that has undergone the smoothing process is matched with the standard voice by the determining means 7, and a result 8 is output.

Now, assuming that the median average of 3 points is taken, when the signal sequence is X (→ (n: sample position)), the median averaged signal sequence xM (→ is as follows).

, Z 帷)−” medlan (”(n−1) +
"k)l"(a+1)'JHowever, medlan
I:'+'/+z) represents the median value of xry, zO. According to this, the maximum and minimum values of three consecutive points in the signal string are discarded and the intermediate value is used as the value at that position, so the original audio waveform is affected by noise and has a diagonal shape. Even when superimposed, extreme changes are absorbed and the continuity of the spectral shape of the original sound wave is maintained. Furthermore, even if the main smoothing is performed, the original shape is preserved for continuous changes, such as a speech waveform in the absence of noise.

Figure 2 shows the concept of median average.
al, at a2°a3, the maximum value a3 and the minimum value a
1 is discarded, a2 is extracted, and 3 points I)t
, l) +, bs, b3 is extracted. FIGS. 3A and 3B show the state before and after smoothing by median averaging.

Effects of the Invention The present invention has the above configuration, and even if noise is superimposed on the audio signal, the noise component can be removed without damaging the original spectral shape of the audio signal. This has the advantage of being able to prevent deterioration in recognition performance due to noise.

In addition, although the above-mentioned example shows an example of smoothing using the median average of three points, the present invention is not limited to the median average of three points.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the speech analysis section of the speech recognition method in an embodiment of the present invention, FIG. 2 is a conceptual diagram of the median average in the embodiment, and FIGS. 3A and B are respectively in the embodiment of the present invention. It is a figure which shows the state before and after smoothing when there is no noise and when noise is superimposed. 1°°° amplifier circuit, 2...filter bank, 20. ,
202°...20n...Band filter, 21. ．． 2
1□,...217L... Rectifier circuit, 221.222
1...22n...integrator circuit, 23. ．． 23. ,. ...23n...Output, 3...Multiplexer, 4
...A/D converter, 5. Smoothing means, 6. Smoothed signal, 7. Discrimination means, 9. Microphone.

Claims (1)

[Claims] A speech recognition method characterized in that speech recognition is performed using a signal obtained by spectrum-analyzing a speech signal, taking the median average of the analysis output signal, and smoothing the signal.