JPS61262798A - Voice section detector - 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 segment detection device for accurately extracting speech segments from an input signal on which noise is superimposed in a speech recognition device.

(Prior Art) A particular problem arises in a voice response recognition system that is composed of a voice response device, a voice recognition device, and a line control device, and provides services in the form of voice interaction to a large number of users through telephone lines. One of the important factors is accurate segmentation of speech sections in speech recognition. The input signal of a speech recognition device includes various ambient noises mixed in from around the speaker, line noise, etc. superimposed on the speech signal uttered by the speaker, and these are constantly changing. .

Several methods have been proposed for accurately extracting voice sections under such circumstances (for example, Japanese Patent Application Laid-Open No. 1986-13).
(See Publication No. 0395). Among these, the characteristic parameters of the input signal containing only various noises at the time when the speaker is not speaking are determined in advance, and the difference from the characteristic parameters of the input audio signal due to the speaker's utterance is also determined. ,
If the difference is greater than or equal to a predetermined threshold and satisfies a predetermined duration, it is defined as the start of the audio, and if the difference is below the threshold and the predetermined duration is satisfied, it is defined as the audio end, and the audio determined by this determination is defined as the audio end. There is a method in which the period from the beginning to the end of the audio is defined as the audio section. A speech segment detection device in a conventional speech recognition device to which this method is applied will be described below with reference to FIG.

In this example, the method for finding the difference between the noise feature parameter and the input audio signal feature/remeter is as follows:
Euclidean distance, city value distance, etc. are common. In the following explanation, a case will be described in which a difference is calculated using Euclidean distance as a distance value from noise.

FIG. 4 is a block diagram showing a speech recognition device including a conventional speech section detection device. In the same figure, 10 is the feature number J
? 11 is a voice section detection device section, and 12 is a recognition section. The speech section detection device 11 includes a noise distance calculation section 11a and a speech section detection section 11b.

The feature parameter creation unit 10 performs frequency analysis on the input audio signal using a pantone matrix filter group with n channels, etc., thereby expressing it as an n-dimensional feature/grammeter time series on the frequency axis.

First, by extracting in advance the characteristic parameters of the input signal at the time when the speaker is not speaking, we can obtain the characteristics and parameters of the noise that does not include the uttered voice (hereinafter referred to as noise/main turn), and NPAT (CH ) [CH=1~
n]) is obtained. Once the characteristic parameters of the noise have been obtained, the time FR of the speaker's utterance (FR is a frame, usually one frame number of m5ec to several tens of meters)
Find the characteristic parameters of the input audio signal in 5ec) The characteristic parameter time series of this input audio signal is hereinafter referred to as the input/mother turn and expressed as INPAT (FR, CH).

The inputs, turns, and noise patterns obtained by the feature noclameter creation section 10 are sent to the noise distance calculation section 11a of the speech section detection device 11.

The noise distance calculating section 11a calculates the difference between the input signal and the noise pattern. In other words, since the input 4-turn is considered to naturally include the above-mentioned noise components in addition to the audio signal, for example, by finding the Euclidean distance between the input A' turn and the noise pattern, we can distinguish between the input audio signal and the input audio signal. Only the audio signal component obtained by subtracting the noise component for each frequency component can be obtained. Specifically, the Euclidean distance (hereinafter referred to as noise distance) at time FB between the noise and the input signal is determined.

Noise distance NDI calculated by the noise distance calculation unit 11a
ST(FR) is sent to the voice section detection section 11b.

The voice section detection unit 11b calculates the noise distance NDIST (PR
) to detect the audio section of the input audio. That is.

When the noise distance NDIST (FR) is compared with a predetermined darkness value and the dog state continues for a predetermined period of time or more.

After determining the start point, noise distance NDIST (F
When R) continues to be small compared to the threshold value for a predetermined period of time or more, it is determined that the audio has ended.

The period from the voice start point to the voice end determined by this determination is detected as a voice section.

A speech section signal indicating the speech section detected by the speech section detection section 11b is sent to the recognition section 12 which performs speech recognition.

(Problems to be Solved by the Invention) However, the conventional voice section detection device having the above configuration has the following problems.

Regardless of the method described above, either the method of calculating the difference using the Euclidean distance to determine the voice interval, or the method of calculating the difference using the city value distance to determine the voice interval, the turn is fixed by noise noise. ing. However, in reality, noise has various fluctuation factors, so it is difficult to say that a fixed noise pattern represents all the noise at each point in time. It is difficult to extract only signal components.

Therefore, noise components that cannot be removed by the fixed noise pattern difference are output as noise distances. Therefore, according to conventional speech recognition devices, among the noise mixed in from the surroundings of a single speaker, steady noise such as the sound of a blower can be removed by difference in noise patterns, but from OA products (personal computers, etc.) There is a problem in that it is impossible to remove the signal sound (bleep sound) that occurs, the click sound that occurs within the line, and such noise is also detected as a voice section.

The present invention eliminates the drawback of the prior art in that the fixed noise filter described above detects incorrect speech sections due to noise components that cannot be removed, and detects speech sections stably and accurately. The purpose of the present invention is to provide an excellent voice section detection device that can perform the following.

(Means for Solving the Problem) The present invention relates to a voice section detection device that analyzes an input voice signal for each frequency channel for each frame, and extracts a voice section based on the analysis result. In order to solve the problems of the prior art, the present invention is configured to include a difference detection means, a sum of squares calculation means, a counting means, a coefficient creation means, a multiplication means, and a speech interval determination means.

The difference detecting means detects the difference between the input audio pattern and the noise pattern obtained in advance in the silent section for each channel. The sum of squares calculation means squares each difference detected by the difference detection means for each next channel, calculates the sum of the sums, and uses the sum as the distance between the input audio pattern and the noise/4' turn in each frame. The counting means compares each difference detected by the difference detection means with a first threshold, and counts the number of channels in which the former exceeds the latter.

The coefficient creation means creates coefficients for the corresponding frame based on the count value counted by the counting means. The multiplication means multiplies the sum of squares obtained by the sum of squares calculation means by the coefficient created by the count creation means to calculate the corrected distance for each of the seven frames. The voice section determination means compares the corrected distance obtained by the multiplication means with the second threshold value, and defines the start point when the state in which the corrected distance is longer than the second threshold value continues for a predetermined period of time, and after determining the start point, A voice section is determined with the end point being the time when the state in which the corrected distance is smaller than the second threshold continues for the first time for a predetermined period of time or more.

(Operation) The difference detection means detects the difference between the input audio output and the noise pattern for each frame and each frequency channel, and supplies the results to the square sum calculation means and the counting means. The sum of squares calculation means calculates the distance between the input speech pattern and the noise pattern based on the information from the difference detection means. On the other hand, the counting means counts the number of channels in which the difference exceeds the first threshold value, and this count is based on the entire noise fluctuation in the frequency direction. The coefficient creation means creates a coefficient taking this amount of variation into consideration, and the multiplication means multiplies the distance obtained by the sum of squares calculation means by the coefficient to obtain a corrected distance. This corrected distance must take into account noise fluctuations, and the speech section is cut out based on the corrected distance. Therefore, the extracted speech section becomes accurate with noise fluctuations removed, and the problems of the prior art described above are solved.

(Embodiment) A voice section detection device according to an embodiment of the present invention will be described in detail below.

FIG. 1 is a block diagram showing the configuration of the voice section detection device of this embodiment. This speech section detection device consists of a noise distance calculation section 1 and a speech section detection section 2.

The noise distance calculation section 1 includes a difference detection section 3, a counting section 4, a coefficient section 5, a sum of squares calculation section 6, and a multiplication section 7. FIG. 2 shows the input waveform of the voice section detection device and the output waveform of each part. In this embodiment, the distance value between the input audio signal and the noise is n
A method of expressing Euclidean distance of dimensional feature parameters is used.

The input signal INPAT (FR, CH) is input to the difference detection section 3 of the noise distance calculation section 1, and the output of the difference detection section 3 is connected to the counting section 4 and the sum of squares calculation section 6.

The output of the counting section 4 is connected to the coefficient section 5, and the sum of squares calculation section 6
and the output of the coefficient section 5 are input to the multiplication section 7.

The output of the multiplication section 7 is input to the speech section detection section 2.

When talking about the operation, first of all, the noise and turn NFAT
(OH) and input cover turn INFAT (FR, CH)
is input to the difference detection section 3. The differential output for each channel is DEL (CH) = INPAT (FR, CH)
-NPAT(CH), n differences are obtained for CH=1 to n.

Each difference output DEL (CH) is input to the counting section 4 and the sum of squares calculating section 6. The counting unit 4 outputs a differential output DEL (CH)
7. A comparison is made for each channel with the threshold NLE"!, the number N(FR)i of channels for which the difference output value is larger is calculated, and the number N(FR) is sent to the coefficient section 5. The number N(FR
) indicates the amount of variation in noise in the frequency direction. 2
The sum calculation unit 6 squares each difference output and calculates the sum thereof. Therefore, the output of the sum of squares calculating section 6 is ND(FR)
=ΣDEL(CH) and CH=. The coefficient part 5 is based on the number N(FR) i based on the calculation formula C(F
R)=fCN(FR)) (f: coefficient function), calculate all the coefficients C(FR) and output it. The multiplier 7 multiplies the sum of squares ND(FR) and the coefficient C(FR) to obtain a noise distance NDIST(FR) and outputs it. In other words, the noise distance is NDIST(FR) = C(FR) x
It is expressed as ND(FR). Here, the coefficient function f is made to have nonlinear characteristics as shown in FIG.

In this way, when the number N (FR) is small and the input audio signal resembles noise, the coefficient C (
FR) is a small value, and the noise distance NDIST(F
R) is output small.

Noise distance NDIST (
FR) is input to the speech section detection section 2, and the speech section detection section 2 compares the noise distance NDIST (PR) with a preset speech section detection threshold VLEV, and determines whether the former is larger than the latter. If the state continues for a certain period of time or more, an ONi signal is output, and if the former continues to be lower than the latter for a certain period of time or more, an OFF signal is output. Then, a voice section is detected using the start point and end point of the ON state as the start and end points of the voice section, respectively.

(Effects of the Invention) As explained above, according to the present invention, the noise pattern obtained in advance in the silent section and the input noise can be combined with the noise pattern obtained in advance in the silent section. The number of channels in which the difference between each channel of the turn exceeds a predetermined threshold is counted, and a coefficient 1 corresponding to the counted number is input.For example, when the count number is small, a coefficient having a small value is input. By multiplying the distance (noise distance) by the distance (noise distance) to find the corrected distance and detecting the voice section based on that, noise, especially single-wavelength noise with a strong peak at a specific frequency (
・It becomes possible to remove fluctuations in the signal sound emitted from Guncon, etc.). Therefore, it is possible to realize stable and accurate detection of voice sections.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the speech interval detection device according to the present invention, FIG. 2 is an input waveform and output waveform of each part of the device of the above embodiment, and FIG. 3 is a diagram showing an example of a coefficient function. FIG. 4 is a block diagram showing a speech recognition device including a conventional speech section detection device. 1... Noise distance calculation unit, 2... Voice section detection unit. 3... Difference detection section, 4... Counting section, 5... Coefficient section, 6... Sum of squares calculation section, 7... Multiplication section.

Claims (1)

[Claims] A voice section detection device that analyzes an input voice signal frame by frame and frequency channel, and extracts a voice section based on the analysis results, includes an input voice pattern and a voice section obtained in advance from a silent section. a difference detection means for detecting the difference between the input audio pattern and the noise pattern for each channel; and a difference detection means for detecting the difference between the input audio pattern and the noise pattern in each frame by squaring each difference detected by the difference detection means and calculating their sum. a sum of squares calculation means for calculating the distance between the two, and a counting means for comparing each difference detected by the difference detection means with a first threshold value and counting the number of channels for which the difference exceeds the first threshold value. and coefficient creation means for creating a coefficient based on the count value by the counting means, and calculating the corrected distance in each frame by multiplying the output of the sum of squares calculation means and the coefficient created by the count creation means. A multiplication means compares the corrected distance obtained by the multiplication means with a second threshold value, and defines the start point when the state in which the corrected distance is greater than the second threshold value continues for a predetermined time or more, and after determining the start point, A speech section detection device comprising: a speech section determining means that determines a speech section with a termination point when a state in which the corrected distance is smaller than a second threshold continues for the first time for a predetermined period of time or more.