JPH03114100A - Voice section detecting device - Google Patents

Abstract

PURPOSE:To detect even a voice of a frictional sound by detecting a start point and an end point of a voiced section of an input voice by a dynamic feature of a voice spectral dispersion and energy. CONSTITUTION:The device is provided with an energy extracting part 20 for extracting the energy setting digital voice data in a prescribed section as one frame, an energy threshold calculating part 30 for adjusting an average value of background noise energy of a frame as an energy threshold, a spectral dispersion extracting part 61 for calculating a spectral dispersion by deriving an average value of a spectrum of the frame by a frequency of the frame, and a spectral dispersion threshold calculating part 71 for adjusting a spectral dispersion average value of a background noise as a spectral dispersion threshold. In this state, the energy and the spectral dispersion of each frame are compared with the energy threshold and the spectral dispersion threshold and whether it is a start point of a voice section or an end point is checked. In such a way, a voice start point of weak energy such as a frictional sound can be detected.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application This invention relates to audio spectral dispersion.
This is a speech section detection device used in a speech recognition system, which detects the start and end points of a voiced section of input speech based on dynamic features of lance ) and energy.

BACKGROUND OF THE INVENTION FIG. 2 is a block diagram showing conventional speech section detection. In the figure, IO is an audio data input means that collects analog audio signals using devices such as a microphone and an analog/digital converter (A/D converter), converts them into digital audio data, and stores the data in a buffer. . 20 is a means for squaring the pattern of each sample for each frame, taking the logarithm, and extracting the energy and the residual energy within this frame. 30
is the background noise of some frames.
d nolse ) and uses the average value as an energy threshold for voice detection. 40 compares the energy in the frame of the energy extraction means 20 with the energy threshold of the energy threshold acquisition means 30, and if the energy in the frame is greater than the energy threshold, this frame is considered to be a voiced section. . If the starting point of the voiced section has not yet been found, this is determined to be the starting point of the voiced section, and if not, the voice starting point detecting means considers it to be continuous speech within the voiced section. 50 detects the voice start point and then converts the energy in the frame of the energy extraction means 20 into the energy threshold acquisition means 30.
If the energy of a consecutive frame is less than the energy threshold, the speech endpoint detection means considers this frame to be the endpoint of speech.

The general voice section detection shown in FIG.
) shows a similar method as above. Furthermore, the ``voice section signal detection circuit'' disclosed in Japanese Patent Publication No. Sho G1-3440 is also a device completed using a similar method as described above based on the zero-cross rate.

Problems to be Solved by the Invention The above-mentioned conventional speech section detection detects speech sections only based on the energy characteristics of the speech. However, in languages such as Chinese, the energy at the beginning of certain speech sounds is very weak and resembles the energy contained in noise. For example, I: (f)% to (S), < (q)N”
If it is a fricative such as ; (c), conventional methods will incorrectly judge it as noise and cannot correctly detect the vocal section of the fricative. Accordingly, the speech sections are also disturbed, and the recognition rate of the recognition device is of course significantly reduced.

In view of this, an object of the present invention is to provide a speech section detection device that can detect sounds such as fricatives that cannot be detected only from conventional energy characteristics.

Means for Solving the Problems In order to solve the above-mentioned problems, the present invention provides an energy extraction unit that extracts energy in i frames of digital audio data within a certain section from an input audio signal; Average the background noise energy of several frames and use the average value as the energy threshold,
There is an energy threshold calculation section that adjusts the energy threshold according to changes in the environment, a spectral dispersion extraction section that first calculates the average value of the spectrum of the relevant frame according to the frame frequency, and calculates the spectral dispersion as a parameter. A spectral dispersion threshold calculation unit that appropriately adjusts the spectral dispersion threshold according to changes in the environment by setting the spectral dispersion average value of the background noise of the frame as the spectral dispersion threshold, and calculates the energy and spectrum of each frame of input audio. Check whether this frame is the start of a speech interval by comparing the variance with an energy threshold and a spectral dispersion threshold, and compare the energy of each frame of the input speech with the energy threshold. The speech section detection device is characterized by comprising a speech section detection section that checks whether or not this frame is the end point of the speech section.

Function: With the above-described configuration, the present invention can detect even voice starting points with low energy such as fricatives using the spectral dispersion method, so that voice sections can be reliably detected.

Example FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, parts that perform the same operations as those in FIG. 2 are given the same reference numerals. 1o is an audio data input section;
20 is an energy extraction section, and 30 is an energy threshold calculation section. The above is the same operation as in FIG. 51 performs linear predictive analysis (L P CN Line
A spectrum calculation unit 61 calculates a spectrum using a spectral average value and a spectral dispersion (Spectrum Coding) function.
This is a spectral variance extraction unit that calculates trum Variance). 90 is the energy of each frame extracted by the energy extraction unit 20 and the spectral variance extraction unit 6
This is a buffer memory that stores the spectral dispersion of each frame extracted in step 1.

The sampling rate of this example
) is, for example, l0KEZ, each frame has 256 points, the overlap between frames is 128 points, and the spectrum is calculated using a linear predictive analysis function for each frame. Spectral calculations are based on sampling and
Space (sampHng 5pace) 0-5
K112 is divided into 16 parts, and each part is one frequency channel. S (n, m) is the spectrum size of the mth frequency channel of the nth frame. Obtain the spectral average value of the frame using the following formula.

Here, 5(n) is the spectral average value of the nth frame. The spectral dispersion of this frame can be calculated by subtracting the spectral average value from the spectrum of each frequency channel within the frame and summing the squares.

V(n) is the spectral dispersion of the 0th frame.

FIG. 3 shows the spectra of four long speech signals /d/, /s/, /a/ and /5llence/ (silence). For these four voices, the spectrum of silence is smoother than the other spectra, so the spectral variance of silence is smaller than the spectral variance of the other three voices. This property allows the sound to be separated from the background signal.

Reference numeral 71 denotes a spectral dispersion threshold calculation unit which calculates a spectral dispersion threshold value according to the following formula based on the spectral dispersion of several frames of audio that has started to be input.

VTR=(ΣV(n))*1. 5 VT here
H is the spectral dispersion threshold based on 10 frames and V(n) is the spectral dispersion of the nth frame within the 5ilence period.

Reference numeral 81 is a voice section detection unit that detects a voice start point and a voice end point. The energy extracted by the energy extraction section 2o and the spectral dispersion extracted by the spectral dispersion extraction section 61 are subjected to spectral dispersion by an energy threshold calculation section 30 and a spectral dispersion threshold calculation section 71, respectively. Compare with a threshold to check whether the frame is the start of speech. That is, as shown in the flowchart in Figure 4, first input the energy of one frame, determine whether the energy is larger than the energy threshold, and if it is smaller than the energy threshold, this Considering the frame as background noise,
Adjust energy threshold.

If not, check if the 5 frames after this frame are greater than the energy threshold,
If it is large, this frame is temporarily set as the voice start point. Then, it is checked whether the spectral dispersion of each frame before this frame is greater than the spectral dispersion threshold until the spectral dispersion of a certain frame becomes smaller than the spectral dispersion threshold. Then, the frame immediately after the frame in which the spectral dispersion becomes smaller than the spectral dispersion threshold is set as the starting point of the voice section. The real voice starting point is this frame after the frame in which the spectral dispersion is less than the spectral dispersion threshold.

FIG. 5 is a diagram showing an example of audio data. In this case, first, five consecutive frames exceed the energy threshold.
Temporarily set the frame as the audio start point. Then this nth
The spectral dispersion value for the frame before the frame is determined and compared with the spectral dispersion threshold value. As a result, the spectral dispersion value is larger than the spectral dispersion threshold from the n-th frame to the n-2nd frame, and since the spectral dispersion value is smaller than the spectral dispersion threshold in the n-3rd frame, the n-2nd frame Let be the starting point of the voice section.

Regarding the detection of the end point of the voice, the energy in the frame of the energy extraction unit 20 is compared with an energy threshold,
If the energy of five consecutive frames is less than some energy threshold, this frame is considered to be the end of the speech. For example, in the case of Chinese, the end point of all speech is a vowel, so the end point of a speech section can be detected accurately by determining it based on energy.

In FIG. 5, since the energy of five consecutive frames from the mth frame to the m+5th frame is smaller than the energy threshold, the mth frame is set as the end point of the voice section.

In each part of the embodiments of the present invention, the start and end points of the voice can be accurately detected based on the characteristics and energy of the spectral dispersion of the voice. It is possible to reduce unnecessary calculations when performing speech recognition, and the recognition rate is also improved. For example, when this speech interval detection device was applied to a speech recognition device for a specific speaker and 100 words in Chinese and city names were recognized, the recognition rate improved from 92% to 98%, and the recognition time was reduced. A significant reduction was also achieved.

This invention is not limited to the above-mentioned embodiments, and can be implemented with appropriate changes as long as the gist is not changed. For example, the method of acquiring a feature called frequency is not limited to a linear predictive analysis spectrum, but can also be applied to a feature representing a linear analysis spectrum or voice frequency.

Furthermore, the data input method is not limited to the microphone, and audio data can also be stored in the memory using a recording device.

In addition, in this example, energy comparison is performed for consecutive 5
This is done for one frame, but the number of frames can be changed depending on the peripheral equipment. Furthermore, the spectral dispersion threshold VT shown in this example
The formula for R is an experimental value, and better results can be obtained by adjusting each coefficient according to the characteristics of the device.

Effects of the Invention The present invention can effectively detect voice sections based on the characteristics of voice energy and spectral dispersion. and,
The practical effects are significant because the recognition rate can be increased and the recognition time can be significantly shortened.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a voice section detection device according to an embodiment of the present invention, FIG. 2 is a block diagram showing the configuration of a conventional voice zone detection device, and FIG. FIG. 4 is an explanatory diagram showing a spectrum, FIG. 4 is a flowchart illustrating voice start point detection, and FIG. 5 is an explanatory diagram showing an example of voice section detection. IO...Audio data input unit, 2o...Energy extraction unit, 30...Energy threshold calculation unit, 41...
・Buffer memory, 51...spectrum calculation section Gl・
... Spectral dispersion calculation section, 71 ... Spectral dispersion threshold calculation section, 81 ...
Voice section detection unit. Figure 4 Figure 11-] Figure 4

Claims (1)

[Claims] An energy extraction unit extracts the energy of digital audio data within a certain interval from the input audio signal as one frame, and an energy extraction unit that extracts the energy in that frame, and an energy extraction unit that averages the background noise energy of several frames and uses the average value as an energy threshold. , an energy threshold calculation section that adjusts the energy threshold according to changes in the environment, a spectral dispersion extraction section that first calculates the average value of the spectrum of the relevant frame according to the frame frequency, and calculates the spectral dispersion as a parameter. A spectral dispersion threshold calculation unit that uses the average value of the spectral dispersion of the background noise of the frame as the spectral dispersion threshold and adjusts the spectral dispersion threshold appropriately according to changes in the environment, and calculates the energy and Check whether this frame is the beginning of a speech interval by comparing the spectral dispersion with an energy threshold and a spectral dispersion threshold, and compare the energy of each frame of input speech with the energy threshold. 1. A speech section detection device comprising: a speech section detection section that checks whether or not the frame of the lever is the end point of the speech section.