JPH0567039B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION <Technical Field> The present invention relates to a voice passing filter that removes acoustic noise by passing only voice.

<Prior Art> When separating speech and noise, until now, noise has been reduced by detecting only a certain type of noise, such as white noise or pulsed noise, and suppressing or removing it.

However, since there are an infinite number of types of noise, conventional methods of suppressing or removing each noise cannot deal with all noises.

<Purpose> The present invention has been made in view of such conventional problems, and its purpose is to effectively remove various types of noise by detecting voice and passing only that voice. The purpose of the present invention is to provide a voice passing filter that can perform the following functions.

<Example> The present invention focuses on the fact that Japanese speech has a basic structure of vowel and consonant pairs, and first detects a vowel interval, and then detects a vowel interval and a predetermined length interval before and after that interval with a high probability of being a consonant interval. In other words, only audio is allowed to pass through.

Further, the above-mentioned vowel-like interval is determined by the following three conditions.

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

Section with small spectral changes (speech stationary part) Area with high power.

The speech passing filter of the present invention detects a vowel-like section based on these three conditions and allows only this section to pass, and will be explained in more detail below with reference to the drawings.

The figure is a block diagram of a voice passing filter according to the present invention. In the figure, 1 is a speech analysis unit that calculates the power P and spectrum y, 2 is a memory that stores the spectrum of the Japanese five-vowel standard pattern, and 3 is the spectrum y output from the speech analysis unit 1 and the memory 2. 4 is a spectral change calculating section that calculates and outputs the spectrum change y'; 5 is a spectral change calculating section that calculates and outputs the spectrum change y'; and 5 delays the audio waveform by a fixed time T frames and sends it to the control section 7. The delay unit 6 is a determining unit that determines a vowel-like section from the power P, the spectrum change y', and the matching distance d with respect to the standard pattern.

In other words, the audio signal is processed by the audio splitter 1.
Sampling is performed at 16KHz (the sampling value at time t is indicated by S(t)), a Hanning window of 16 ms is applied, and the power P, which is the sum of squares of the sampling values at every 8 ms frame period, is obtained. The spectrum y is obtained by Fourier transform within the frame (the power and spectrum of the t-th frame are denoted by P(t) and y(t)). The spectrum y(t) thus obtained is sent to the matching section 3, and the power P(t) is sent to the determining section 6.

The matching unit 3 compares the spectrum y(t) at each time t with the spectrum of the Japanese five-vowel standard pattern stored in the memory 2, and calculates the Euclidean distance (matching distance) d(i, t) between the spectra.
(However, i=1, 2, 3, 4, 5 are vowels/
a/, /i/, /u/, /e/, /o/) and outputs it to the determination unit 6. Furthermore, the Euclidean distance between spectra separated by several frames is determined in the spectrum change calculation section 4 and sent to the determination section 6 as the spectrum change y'(t).

When the matching distance d(i, t), power P(t), and spectrum change y'(t) are input to the determination unit 6 in this way, the determination unit 6 determines the following conditions 1) and 2).
It is determined whether or not the frame is satisfied, and if it is, it is determined that the frame is within the vowel interval and the vowel flag is set.
VF(t)=1, and if not satisfied, vowel flag VF(t)
=0 is output to the control unit 7, respectively.

Condition 1 Power P(t) is greater than threshold a1,
In addition, the spectral change y'(t) is smaller than the threshold a2, and the minimum value min d(i, t) [i=1-5] of the matching distance d(i, t) is smaller than the threshold a3.

Condition 2 Spectrum y'(t) is larger than threshold a4, and matching distance d(2,t) or d(3,
t) is smaller than the threshold a5. (Taking into consideration the devoicing of vowels /i/, /u/ that follow unvoiced consonants) Then, the control unit 7 sets If more than the number of frames exist, O(t)=S(t-T) is output, and if not, O(t)=0 is output.

In this manner, the voice passing filter of the present invention detects a vowel section and a predetermined length section before and after the vowel section, that is, a vowel-like section, from the input sound, and passes only this section, and only the voice passes. , many types of acoustic noise can be effectively reduced.

<Effects> As explained in detail above, the speech passing filter of the present invention includes a speech analysis section that obtains the power and spectrum of an input sound, and a spectrum of a standard vowel pattern stored in advance and output from the speech analysis section. a matching unit that calculates a matching distance by comparing the spectrum with the spectrum output from the audio analysis unit; a spectrum change calculation unit that calculates and outputs a spectrum change between spectra that are several frames apart from the spectrum output from the audio analysis unit; a first condition that the spectral change is greater than a first threshold value, the spectral change is less than a second threshold value, and the minimum value of the matching distance is less than a third threshold; a determination unit that determines whether a second condition is satisfied that the change is greater than a fourth threshold and the matching distance is less than a fifth threshold; and a determination unit that responds to the output of the determination unit; , if the first or second condition is satisfied,
A control unit that determines that the input sound is speech and outputs the input sound detects a vowel-like section from continuous input sounds and outputs only the input sound in the section as speech. It is a thing,
The first and second conditions are set in order to detect vowel-like intervals, and the method detects vowel-like intervals by taking into consideration the devoicing of the vowel following a voiceless consonant. remove acoustic noise,
It becomes possible to reliably output only audio. Furthermore, by installing the audio passing filter of the present invention in various devices, the following effects can be achieved.

Personal radio: Normally, if the input gain is increased and used, there is a risk that the noise will be amplified to a small level and cause discomfort, but if the gain is increased after passing through the voice passing filter of the present invention, only the voice will be emphasized. It becomes very easy to hear.

Telephone device: The voice passing filter of the present invention does not allow noise from airplanes, cars, etc. to pass through, so by using it in a telephone, it is possible to clearly transmit the voice of a telephone from an airport or near a road.

Tape recorder: Even if the content is recorded in a noisy environment, it can cut out the noise and clearly output only the audio when playing back.

Speech recognition device: If noise enters while waiting for voice input, the recognition device will operate incorrectly as voice input. In this case, recognition results due to noise can be rejected from the recognition score, but even if speech comes in during this time, it will not be accepted. However, by using the filter of the present invention, malfunctions of the recognition device due to noise can be avoided, and mistakes in receiving voice input can also be prevented.

[Brief explanation of the drawing]

The figure is a block diagram of a voice passing filter according to the present invention. 1 is a speech analysis section, 2 is a vowel standard pattern memory, 3 is a matching section, 4 is a spectrum change calculation section, 5 is a delay section, 6 is a determination section, and 7 is a control section.

Claims (1)

[Scope of Claims] 1. A voice analysis unit that calculates the power and spectrum of an input sound, and a matching distance is determined by comparing the spectrum of a standard vowel pattern stored in advance with the spectrum output from the voice analysis unit. a matching section; a spectral change calculation section that calculates and outputs a spectral change between spectra several frames apart from the spectrum output from the speech analysis section; a first condition that the change is less than a second threshold and the minimum value of the matching distance is less than a third threshold, or the spectral change is greater than a fourth threshold, and a determination unit that determines whether or not a second condition that the matching distance is smaller than a fifth threshold is satisfied; and a determination unit that determines whether or not the first or second condition is satisfied in response to an output of the determination unit. and a control unit that determines that the input sound is voice and outputs the input sound when the input sound is voice. filter.