JPS61233791A - Voice section detection system for voice recognition equipment - 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 The present invention relates to a speech segment detection method in a speech recognition device.

In conventional speech recognition devices, when detecting a speech section, the speech section is detected using a threshold value such as speech power information, but if a certain threshold value is used due to surrounding noise etc. Section detection becomes difficult, and in the case of a variable threshold,
The threshold value increased due to the noise level, and it was not possible to detect speech sections buried in noise.

FIG. 3 is a diagram for explaining an example of a conventional voice section detection method, and FIG. 3(A) shows a case where a fixed threshold is used.

FIG. 3(B) shows the case of variable threshold value;

(G) is the voice power, (b) is the threshold value, and (c) is the voice interval signal. In the case of a variable threshold value, (B) shows the threshold value within the range of b' as shown in (b) of the figure. It is variable. Therefore, in the conventional voice section detection described above, when detecting a sound section based on voice power information, etc., if a certain threshold value is used as shown in Figure (A), it becomes impossible to detect that section. Furthermore, in the case of a variable threshold value as shown in FIG. 12(B), the threshold value increases in accordance with the noise level, and useful information in that section may be missing. However, useful information is still contained near the beginning and end of the voice, and it has been desired to extract this information.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to enable a speech recognition device to stably detect speech sections even under noisy conditions.

In order to achieve the above object, the present invention includes a bandpass filter for analyzing input audio, a means for obtaining an audio power signal that is the sum of each channel of the bandpass filter, and
means for obtaining frame feature data from the input spectrum analysis data of each channel; means for obtaining a voice section by comparing the voice power signal with a certain threshold; and a means for obtaining a voice section from the obtained voice section signal. When there is a silent section, when the start of speech is detected in a speech recognition device that has means for generating a one-word signal that uses the silent section as a word delimiter, the data one frame before the first frame is used as voice information. If the inter-frame distance is smaller than a certain value, you can use that as the first frame and perform it backwards in a certain frame section, or after detecting the start of the audio, cut out the audio section and start the next frame at the end. The characteristic data is checked, and if the inter-frame distance from the end frame is less than a certain value, it is regarded as the true end and the process is continued until there is no more feature data left. Hereinafter, the present invention will be explained based on examples.

Figure 1 is an electrical circuit diagram for explaining an embodiment of the present invention using the BTSP method. Part, 5
is a power information section, 6 is a section detection section, and 7 is a section correction section.The BTSP system uses the formant information of the voice by binarizing it for each frame period using a least-diamond curve.

Hereinafter, the present invention will be explained in detail with reference to FIG. 2. In FIG. 2, (q) is the audio power, (b)
is the threshold, (c) is the voice section signal, D is the start end, E is the end,
(d') is audio data (BTSP).

CB) is the interframe distance, and (f) is the true voice section. Now, it is assumed that the threshold value changes depending on the surrounding noise level, etc., and if the threshold value at that time is detected as shown in the figure,
The inter-frame distance between the detected starting edges D(0) and D(-1) is calculated, and if the inter-frame distance is less than or equal to a certain value, that frame is determined to be the true starting edge. In other words, the beginning contains something with relatively weak voice power, such as a consonant, and it is likely that this will be drowned out by noise, but in this case,
Since the correlation with frame 0 is strong, the distance between the frames becomes small in that sense. Focusing on this, the system searches backwards until the inter-frame distance becomes large, defines the period up to this large distance as a voice section, and uses that as the true starting point. Note that since actual noise has a low correlation with speech, its breakpoint is relatively clear. In this way, we can search backwards between a limited number of n frames,
Detect the true starting point. Further, the same process is performed for the termination, and the point where the inter-frame distance becomes large is determined as the true termination. This detection section is also a certain number of n frames from the end, and when there is no feature data within it, the point Q, e is set as the end. In this way, it becomes possible to detect information about the beginning and end of a word that is lost due to noise or the like.

As is clear from the above description, according to the present invention, it is possible to stably detect a speech section even under noise without losing information at the beginning or end of a word.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing an example of an electric circuit used in carrying out the present invention, Fig. 2 is a time chart for explaining an embodiment of the present invention, and Fig. 3 is a diagram of a conventional voice section detection method. FIG. 3 is a signal waveform diagram for explaining an example. ■... Microphone, 2... Preprocessing section, 3... Frequency analysis section, 4... Feature data extraction section, 5... Power information section, 6... Section detection section, 7...・Section correction section. Figure 1 Figure 2

Claims (2)

[Claims] (1) A bandpass filter that analyzes input audio, a means for obtaining an audio power signal that is the sum of each channel of the bandpass filter, and obtaining frame feature data from spectrum analysis data of each input channel. means for obtaining a speech interval by comparing the speech power signal with a certain threshold value; In a speech recognition device having a means for generating a word signal, when the start of speech is detected, if the data one frame before the first frame is voice information and the interframe distance is smaller than a certain value, it is detected. A speech section detection method in a speech recognition device, characterized in that the first frame is set as the first frame, and the detection is performed retroactively to a certain frame section. (2) A bandpass filter for analyzing input audio, a means for obtaining an audio power signal that is the sum of each channel of the bandpass filter, and obtaining frame feature data from spectrum analysis data of each input channel. means for obtaining a speech interval by comparing the speech power signal with a certain threshold value; In a speech recognition device having a means for generating a word signal, after detecting the start of speech, a speech section is cut out, and at the end, subsequent feature data is checked, and the interframe distance from the end frame is determined to be a certain value. 1. A speech section detection method in a speech recognition device, characterized in that if the following is true, the detection is performed until there is no more feature data.