JPH03253899A - Voice section detection system - Google Patents

Abstract

PURPOSE:To reduce misdetection by using plural parameters and detecting a start end when the number of parameters which exceed a threshold value is larger than a reference value. CONSTITUTION:A band power detection part 2 detects each band power value by using a band-dividing device such as a band-pass filter, a total power detection part 3 detects the total power, and a variation calculation part 4 calculates differences between current power and last power, parameter by parameter, and detects variation in each parameter. Then comparison parts 5 - 12 compare them with threshold values found previously by a threshold value generation part 13, parameter by parameter, and a counter 14 counts parameters Pn exceeding the threshold value EPSILONn and a comparison part 15 compares the number C obtained by the counter with the reference value LAMBDA predetermined by a reference value generation part 16; when the number C of the counter exceeds the reference value LAMBDA, the start end of the voice is decided. Consequently, the correct voice section is detected even in the presence of ambient noise.

Description

[Detailed Description of the Invention] Ritsuka Edasaki The present invention relates to a method for detecting speech sections in speech recognition, and is effective for detecting speech sections when applied to speech recognition under noisy conditions, particularly in automobiles, factories, etc. Regarding the method.

Detection of speech sections is a very important issue in realizing a speech recognition device that can be used in a short period of time, and has a great influence on subsequent recognition results. A common method for detecting speech sections is to detect the section from the time when the power of the input signal exceeds a predetermined threshold to the next time when the power of the signal falls below the threshold as a speech section. However, with this method, if there is noise in the surroundings, the voice section may be detected incorrectly.

Figure 3 is a diagram for explaining one method for solving the above-mentioned problems. One possible countermeasure is to use two thresholds T□ and T2 as shown in Figure 3. .

In this method, threshold values T,,T are set in advance.

(Tz > T), and detect the interval S from the time point ○ when the input signal power exceeds T to the next time t2 when it becomes less than T, and This method detects this section S as a voice section only when there is a portion exceeding the second threshold T2.

However, even with this method, it is difficult to detect speech sections when there is noise in the surroundings, and it is particularly difficult to detect sections when the noise level at the beginning and end of the speech is different. There are drawbacks. In addition, noise often exists biased toward a certain band, and there are cases where only the components of a certain band exceed the above-mentioned threshold, but this method has the disadvantage that it does not take measures for such cases. be.

And---Last 3-4 The present invention was made in view of the above-mentioned circumstances, and
In particular, the purpose of this invention is to provide a voice section detection method that can detect a correct voice section even when there is surrounding noise.

In order to achieve the above object, the present invention provides (1) a voice section that collects voice, converts it into an electrical signal, and detects a section where voice exists based on the converted signal; In the detection method,
The electrical signal is frequency-converted and divided into a plurality of frequency bands, and has means for determining the power of each band and the total power of all the bands, and means for determining the amount of change in each band and the total power, Each band power, total power The change in each band power and the change in total power are used as parameters for section detection. (2) The sound is collected and converted into an electrical signal, and the section where the sound is present is detected based on the converted signal. In the speech interval detection method,
The electrical signal is frequency-converted and divided into a plurality of frequency bands, and has means for determining the power of each band and the total power of all the bands, and means for determining the amount of change in each band and the total power, Each band power, total power, change in each band power, and change in total power are used as parameters for section detection, and each parameter for which a predetermined number or more of the plurality of parameters is determined in advance. or (3) collecting the sound and converting it into an electrical signal, and then determining the section where the sound exists based on the converted signal. In the voice section detection method for detecting the electrical signal, the electrical signal is frequency-converted, divided into a plurality of frequency bands, and means for obtaining the power in each band and the total power in all bands, and the change in each band and the total power is calculated. It has a means for determining, and uses each band power, total power, change in each band power, and change in total power as parameters for section detection, and This device is characterized in that when the number exceeds a predetermined threshold value for each parameter, it is determined to be the beginning of the voice, and when it becomes less than the threshold, it is determined to be the end of the voice. Hereinafter, the present invention will be explained based on examples.

As described above, the present invention includes a voice section detection method that collects voice, converts it into an electrical signal, and detects a section where the voice exists based on the converted signal, in which the electrical signal is frequency-converted. , comprising means for dividing the frequency bands into a plurality of frequency bands and calculating the power of each band and the total power of all bands, and means for calculating the change in each band and total power, and the change in each power and each power. minutes is used as a parameter for section detection, and when a predetermined number or more of the plurality of parameters exceeds a predetermined threshold for each parameter, it is determined that it is the beginning of the voice,
Similar to starting edge detection, end detection uses the total power of each band and the change in each band power and total power as detection parameters, compares them with a predetermined threshold, and calculates the number of objects that fall below the threshold. When this number exceeds a predetermined number, it is determined that the terminal has reached the end.

FIG. 1 shows the voice start detection method according to claim 1 of the present invention;
1 is a block diagram for explaining an embodiment of the voice end detection method according to claim 2 and the voice section detection method according to claim 3 when these rain detection methods are combined;
1 is an audio input section, 2 is a band power detection section, 3 is a ]-tal power detection section, 4 is a variation calculation section for each power, 5 to 12
15 is a comparison section, 13 is a threshold value generation section, 14 is a counter, 16 is a reference value generation section, and 17 is a speech recognition section.

First, in the audio input section 1,
An electrical signal converter is used to convert the sound into an electrical signal, a band power detection section 2 detects the power of each band using a band division device such as a band pass filter, and a total power detection section 3 detects the power of each band. In addition to detecting the total power, the change calculation unit 4.

7-- Calculate the difference between the current power and the previous power for each parameter, detect it as a change in each power, and compare it with the threshold previously determined by the threshold generator 13 for each parameter. - Compare each parameter P. is the respective threshold ε. The counter 14 counts the number of objects that exceeds λ, and the comparison unit 15 compares the counter number C with a reference value λ predetermined by the reference value generation unit 16, and the counter number C becomes the reference value λ.
When this happens, it is determined that this is the beginning of the audio. In Figure 1, the band is divided into three bands: low, middle, and high, and in this case, a total of eight parameters are calculated by the power of the low, middle, and high ranges, the total power, and the change in each power. This will be used to detect the start end.

The total power can be determined by simply calculating the sum of the powers of low, middle, and high frequencies, or by using an all-pass filter. As a method for calculating the threshold value, a method is used in which noise-only data is taken in before audio is input, and a threshold value proportional to the data is obtained for each parameter.

Similarly to the starting edge detection, the termination detection according to claim 2 also includes the total power of each band, the total power of each band,
Using the change in total power as a parameter for end detection, comparing it with the threshold value for each parameter determined in advance, each parameter P is determined. A counter counts the number of sounds below a threshold value ε, and when the number C on the counter becomes a predetermined reference value λ or more, it is determined that the audio has come to an end. Note that the threshold value ε used in the start edge detection. Reference value λ and threshold value ε used for end detection. , the base value λ is not necessarily the same.

The speech recognition section 17 is a speech recognition section using a known method, and performs matching between the speech pattern in the speech section detected by the above method and a standard pattern registered in advance, and recognizes the speech input from the speech input section 1. It is something that recognizes.

FIG. 2 is a flowchart for explaining an embodiment of the voice section detection method according to claim 3 of the present invention. ) to detect the start end, and if the start end is found,
From then on, in the # edge detection method (■ part) of claim 2,
Detects the end, and if the end is found, from then on,
The start point is detected again, and if the start point is not found for more than the specified time, it is determined that the audio input has ended.
The section detection is finished, and speech recognition is performed using the detected speech section.

As is clear from the above description, according to the invention of claim 1, if only the voice power is used as a parameter for detecting the start point, the time when the threshold value is calculated,
If the magnitude of the surrounding noise at the time of detecting the start point is different, the start point detection cannot be performed well. However, the start point detection method of the present invention uses not only the voice power but also the change in the voice power, so it is difficult to detect the start edge. The start end can be detected even in such cases. Additionally, when noise is concentrated in a certain band, such as inside a car where most of the noise is in the low range, conventional methods may not be able to detect the start end due to the influence of the components in that certain band. However, the start edge detection method of the present invention uses multiple parameters and detects the start edge when the number of parameters exceeding a threshold value is greater than or equal to the reference value, so it requires one more step than the conventional method. The weight of one parameter becomes smaller, reducing the possibility of false detection.

Further, according to the invention of claim 2, if only the voice power is used as a parameter for detecting the end, the threshold value is determined by the difference between the surrounding noise at the time when the end is calculated and the time when the end is detected. If the sizes are different, the termination cannot be detected successfully, but the termination detection method of the present invention uses not only the voice power but also the change in the voice power, so the termination can be detected even in such a case. In addition, when the noise is concentrated in a certain band, such as in a car where most of the noise is in the low range, conventional methods sometimes fail to detect the end due to the influence of the components in that certain band. The end detection method of the present invention uses a plurality of parameters, and detects the end when the number of parameters that fall below a threshold is greater than or equal to a reference value. The weight of the parameters becomes smaller, and the possibility of false detection is reduced.

Furthermore, according to the invention of claim 3, noise can be reduced by combining the voice start detection method described in claim (1) and the voice end detection method described in claim (2). A method for detecting voice segments in the existing state is realized.

[Brief explanation of drawings]

FIG. 1 is a block diagram for explaining an embodiment of the present invention, FIG. 2 is a flowchart for explaining the operation of the present invention, and FIG. 3 is for explaining an example of a conventional speech interval detection method. This is a diagram for 1... Audio input section, 2... Band power detection section, 3.
. . . Total power detection unit, 4 . Value generation section, 17...Speech recognition section. Figure 1

Claims (1)

[Claims] 1. In a voice section detection method that collects voice and converts it into an electrical signal, and detects a section where the voice exists based on the converted signal, the electrical signal is frequency-converted; It has means for dividing the frequency bands into a plurality of frequency bands and calculating the power of each band and the total power of the whole band, and means for calculating the change in each band and the total power. The change in band power and the change in total power are used as parameters for section detection, and when a predetermined number or more of the plurality of parameters exceeds the predetermined threshold for each parameter, the voice is detected. A voice section detection method characterized in that it is determined that it is a starting point. 2. In the voice section detection method, which collects voice and converts it into an electrical signal, and detects the section where the voice exists based on the converted signal, the electrical signal is frequency-converted and divided into multiple frequency bands. and means for determining the power of each band and the total power of all bands, and the means of calculating the change in each band and total power, and the power of each band, the total power, and the change in power of each band, The change in total power is used as a parameter for section detection, and it is determined that the end of the audio is reached when a predetermined number or more of the plurality of parameters mentioned above falls below a predetermined threshold for each parameter. A voice section detection method characterized by: 3. In the voice section detection method, which collects voice and converts it into an electrical signal, and detects the section where the voice exists based on the converted signal, the electrical signal is frequency-converted and divided into multiple frequency bands. and means for determining the power of each band and the total power of all bands, and the means of calculating the change in each band and total power, and the power of each band, the total power, and the change in power of each band, The change in total power is used as a parameter for section detection, and when a predetermined number or more of the plurality of parameters exceeds a predetermined threshold for each parameter, it is determined that the beginning of the voice is present. , a voice section detection method characterized in that it is determined that the end of the voice is reached when the value becomes less than a threshold value.