KR19990056313A - Speech Segment Detection Method in Speech Recognition System - Google Patents

Abstract

According to the present invention, in a speech recognition system, an energy for each frequency band of an input signal is calculated using a filter bank, and a speech recognition is performed to detect an accurate speech section from the input signal by grasping the correlation of the calculated energies. A method for detecting a speech section in a system, which uses a high pass filter to emphasize a high frequency region of an input signal, subdivides an input signal with a high frequency region into a frame of a predetermined size using a hamming window, After performing the FFT, the energy corresponding to each frequency is obtained by using a triangular filter, and the correlation coefficients are calculated to calculate the voice interval judgment index. In case of abnormality, it detects that the input signal is a voice signal, and in case of less than the threshold value, noise signal It can be detected, even in the case of a signal mixed with a relatively large noise compared to the prior art, it is possible to accurately detect the speech section, and even in the case of consonants that were difficult to detect due to the noise and difficult to detect to some extent It is effective.

Description

Speech Segment Detection Method in Speech Recognition System

The present invention is a speech recognition system for recognizing a human voice, using the filter bank (Filter Bank) to calculate the energy of each frequency band of the input signal, and to determine the correlation of the calculated energy from the input signal The present invention relates to a speech section detection method in a speech recognition system capable of detecting an accurate speech section.

In general, a speech recognition system for recognizing a natural sound such as a human voice includes a speech section detection unit 1 for detecting a speech section from an input signal as shown in FIG. 1 and the speech section using an MFCC coefficient. A feature coefficient extraction unit 2 for extracting the feature from the speech section detected by the detection unit 1, and using the Hidden Markov Model (HMM) and VMS Variable Multi-Section Vector Quantization (VQ) algorithm to recognize the voice signal The speech recognition unit 3, a database 4 storing word model parameters learned by the speech signal, and a word recognized by determining effectiveness of the speech signal recognized by the speech recognition unit 3 It consists of the post-processing part 5 which outputs.

In the speech recognition system configured as described above, detecting the correct speech section from the input signal is a pre-processing part of the speech recognition system, which is a very important task as a precondition for determining the performance of the system.

However, in a noisy external environment, it is difficult to detect an accurate speech section due to noise components.

Accordingly, conventionally, the speech section is detected by comparing the threshold value calculated from the background noise using information such as the maximum energy or average energy of each frame of the input signal, zero crossing rate, or the like.

However, this detection method has a problem that the speech section is relatively well detected when the background noise is small, but the speech section detection performance is remarkably degraded when the background noise is relatively large.

In addition, consonants such as garments, jigs, teats, and kiosks, which are not easily distinguished from noise components, are not included in the voice section at all, and thus there is a problem that cannot be detected.

The present invention has been made to solve the above problems, the purpose of which is to calculate the energy of each frequency band of the input signal using a filter bank, and through the correlation of the calculated energy signal and noise signal SUMMARY OF THE INVENTION An object of the present invention is to provide a speech section detecting method in a speech recognition system capable of effectively detecting only a speech section from an input signal.

The speech segment detection method in the speech recognition system of the present invention for achieving the above object, by using a high-pass filter to emphasize the high frequency region of the input signal, the input signal of the high frequency region is highlighted using a Hamming window After subdividing by frame, FFT by subdivided frame unit, the energy corresponding to each frequency is calculated by using a triangular filter, and the correlation coefficient between each energy is calculated to calculate the voice interval judgment index. After that, if the voice interval determination index is greater than or equal to the threshold value, the input signal is detected to be a voice signal.

1 is a block diagram of a general speech recognition system;

2 is a flowchart for detecting a speech section in a speech recognition system according to the present invention;

3 is a view showing a filter bank according to the present invention.

<Description of the symbols for the main parts of the drawings>

1: Voice section detection unit 2: Feature coefficient extraction unit

3: speech recognition unit 4: database

5: post-processing unit

Hereinafter, a method of detecting a voice interval in a voice recognition system according to the present invention will be described in detail with reference to FIG. 2.

First, the high frequency region is emphasized using the high pass filter of the input signal input to the voice recognition system as shown in Equation 1 below (S10).

Subsequently, in step S10, the input signal in which the high frequency region is emphasized is subdivided into frames of 20 ms size by overlapping by 10 ms using a Hamming window as shown in Equation 2 below (S11).

Then, if a 256-point Fast Fourier Transform (FFT) is performed in each frame unit subdivided in step S11 (S12), the FFT result is a triangular filter as shown in FIG. That is, the energy corresponding to each frequency is obtained using the filter bank (S13).

Subsequently, a correlation coefficient is obtained with respect to the background noise signal prepared in advance for the signal passing through the filter bank in step S13 using Equation 3 (S14).

Thereafter, the voice interval determination index is calculated from the correlation coefficient obtained in the step S14 as shown in Equation 4 (S15), and whether the voice signal is determined is determined using the voice interval determination index (S16).

That is, when the voice interval determination index is greater than or equal to a threshold value, the input signal is detected as a voice signal (S17), and when it is less than or equal to the threshold value, a noise signal is detected (S18).

As described above, the present invention calculates a correlation between energy values of respective frequency bands in a frequency domain, and separates a voice signal from an input signal based on this, thereby resulting in a relatively large noise. Even in the case of mixed signals, the speech section can be accurately detected, and in the case of consonants that were difficult to detect because of difficulty in distinguishing from noise, it is possible to detect to some extent.

Claims (1)

In the method for detecting the speech section in the speech recognition system, A first step of emphasizing a high frequency region of the input signal using a high pass filter, a second step of subdividing the input signal of which the high frequency region is emphasized into a frame having a predetermined size using a hamming window, and the second step A third step of performing an FFT on the input signal divided in step 2 for each frame unit; a fourth step of obtaining an energy corresponding to each frequency using a filter bank from the FFT result performed in the third step; A fifth step of obtaining a correlation coefficient with respect to the background noise signal prepared in advance in the input signal passed through the filter bank in a fourth step; and a sixth step of calculating a voice interval determination index using the correlation coefficient obtained in the fifth step. And when the voice section determination index calculated in the sixth step is equal to or greater than a threshold value, detect that the input signal is a voice signal. Speech period detection method for a speech recognition system which comprises a seventh step of detecting that.