JPS6217800A - Voice section decision 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 Technical Field The present invention relates to a speech segment determination method for separating speech and noise from input sound.

<Prior Art> When separating speech and noise, conventional techniques have been to detect only certain types of noise, such as white noise or pulsed noise, and to suppress the noise in order to reduce the noise.

However, there are an infinite number of types of noise, and conventional methods that change the suppression method for each type of noise cannot deal with all types of noise.

Purpose> The present invention has been made in view of such conventional problems, and
Rather than detecting and suppressing specific noise, the present invention aims to provide a speech interval determination method that can very easily remove a large number of types of noise by separating speech and noise from input sound. It is something. Furthermore,
It is an object of the present invention to provide a speech interval determination method that determines speech intervals based on the presence or absence of vowels and is capable of separating the obtained speech intervals from noise intervals.

<Example> In Japanese speech whose basic structure is a pair of consonants and vowels,
The following four conditions are listed as conditions for a vowel-like interval.

■ Sections with high power.

Φ　A section with small spectral changes (voice stationary part).

■ An area where the matching distance with the standard vowel pattern is small.

■　Sum of absolute values of kegstrum coefficients or larger interval.

The present invention detects vowel intervals based on the conditions ■ and ■ out of these four conditions and separates them from noise intervals, thereby omitting matching with the standard pattern of vowels. When the voice section can be determined using a simpler hardware configuration, Kvf is found.

Next, the system of the present invention will be explained in detail based on the drawings.

FIG. 1 is a block diagram of a speech segment determination device implementing the method of the present invention.

In the figure! 2 is a speech analysis section, 2 is a cepstrum sum calculation section, and 3 is a judgment section.

As shown in the block diagram of FIG. 2, the speech analysis section 1 includes an autocorrelation coefficient calculation section 4, a linear prediction coefficient calculation section 5,
It consists of a cepstrum coefficient calculation section 6 and a power calculation section 7.

The autocorrelation coefficient calculation unit 4 calculates the sampling value 5(i) (where 1≦l≦256) and the analysis order np=24,
Based on the processing flow in Figure 3, the autocorrelation coefficient (Ri)
(However, 1≦i≦np+In is required.

On the other hand, the linear prediction coefficient calculation unit 5 inputs the autocorrelation coefficient (Ri) from the calculation unit 4, and follows the processing flow of FIG. 4 to calculate the linear prediction coefficient A(i) (where ■≦i≦n
p Eight autocorrelation coefficients P(i) and residual power E
Calculate (i). In addition, the cepstral coefficient calculation unit 6 calculates the linear prediction coefficient A(i) (l≦i≦np
), the kegstrum coefficient c(i)(
1≦i≦np).

Furthermore, in the power calculation unit 7, the sampling value 5(i)(
+≦i≦256), the power P is determined based on the following equation.

Next, the operation will be explained.

First, the audio signal is sampled at 16KH2 in the audio analysis unit 1, and the sampling value at time t is 5(t)), and a Hayung window of 16 ms is applied, and the power P and LPC cepstrum are calculated every 8 ms frame period. Find C.

(Note that the power and cepstrum of the first frame are P (t) and c (t), respectively). The obtained LPC kegstrum c (t) is input to the cepstral sum calculation section 2 at the next stage, where it is output as the end of low-order (up to the 24th order) kegstrum coefficients.

The power P (tJ and cepstral sum C
(t) is sent to the determination unit 3 (/c), where the following determination is made.

That is, the power P (t) is larger than the threshold al (
(see Figure 5), and the cepstral sum C (υ is the threshold a2
If it is larger (see FIG. 6), it is determined that the frame is within the vowel interval Vc6. Then, the interval tI<t<t2
(however, t 2 - t + > 84 frames), if the voiced section is 21 frames or more, and the number of frames determined to be vowel sections is 174 or more, which is the length of the voiced section, then the section tI < If t<t2, it is determined that it is a voice section, and if it is less than 1/4, it is determined that it is a noise section.

In this way, it is possible to determine and separate the speech section and the noise section in the input sound based on the relationship between the threshold and the ratio of the vowel diagram 5 length to the input sound section length. A particular feature of this method is that it does not perform matching processing with a standard vowel pattern when detecting vowel sections, so it is possible to determine speech sections with a very simple hardware configuration.

Note that the ratio of the vowel section length to the input sound section length is not limited to the embodiment and can be determined as appropriate.

Effect> As explained in detail above, the speech interval determination method according to the present invention detects a vowel interval in an input sound, calculates the ratio of the detected vowel interval length to the input sound interval length, and calculates the ratio of the detected vowel interval length to the input sound interval length. Since the input sound is determined to be speech when it is larger than the threshold, it is possible to detect only speech from the input sound, and when detecting vowel sections, matching process a with the standard vowel pattern is performed. Therefore, there is a great effect that the configuration of the software can be significantly simplified.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a speech segment determination device adopting the method of the present invention, Fig. 2 is a block diagram of the speech analysis section of Fig. 1, Fig. 3 is a processing flow diagram of the autocorrelation coefficient calculation section, and Fig. 4
Figure 5 is a diagram showing the processing flow in the linear prediction coefficient calculation unit, Figure 5 is a diagram showing the relationship between the power threshold, noise and speech, and Figure 6 is a diagram showing the relationship between the Kegnutrum sum threshold, noise and audio. It is. 1 is a speech analysis unit, 2 is a kegnutrum sum calculation unit, 3 is a determination unit, 5(t) is a sampling value, c(t) is an LPG kegnutrum, P(t) is power, and C(t) is a kegnutrum sum. Agent Patent attorney Aihiko Fuku (and 2 others) Figure 1 91 color subtitles Figure 4 Figure 6

Claims (1)

[Claims] 1. A voice section determination method that detects a vowel section in an input sound, calculates the ratio of the detected vowel section length to the input sound section length, and determines that the input sound is speech when the ratio is greater than a threshold value.