JPS61256394A - Voice section detection system - 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 Technical Field: 1. Field of the Invention The present invention relates to a speech segment detection method in a speech recognition device.

BACKGROUND OF THE INVENTION In general, the power of a voice signal differs depending on the speaker. Therefore, when performing AD conversion of an audio signal, an AGC circuit is provided before the AD converter in order to maintain the effective number of bits. This AGC circuit eliminates the difference in voice power between speakers, but the gain of the AGC circuit becomes large during silent sections, and becomes small during speech sections. Therefore, in the signal after the AGC circuit, the difference in power between a silent section with noise and a voice section is small. Voice sections are detected based on a certain threshold by detecting power, but if the threshold is varied depending on the noise power in silent sections, the power threshold will rise relatively in voice sections due to the difference in the gain of the AGC circuit. , low-power sounds such as consonants following consonants were sometimes missing. For example, “5TOP
In words such as ``F'' and ``Z'', the consonant after a silent period in the word may be missing.

FIG. 3 is a waveform diagram for explaining the shortcomings of the prior art described in "-".
The relationship is shown using "S T OP" as an example (
The shaded area is noise). In this case, the #P II portion is missing, and the voice section signal becomes as shown in Figure (b). That is,
In conventional speech recognition devices, speech intervals are detected using power information of the input signal and a threshold determined based on the noise power during periods when speech is not being uttered. In this case, after the input signal is passed through the AGC circuit, the gain of the AGC circuit is different when there is no utterance and when there is utterance, so the consonant after the consonant may be reduced by the noise power during the unuttered time and not be detected as speech. Ta.

Purpose 1 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.

Purchase - Kazunari.

In order to achieve the above object, the present invention provides an AGC circuit that eliminates differences in input voice power depending on speakers;
A means for obtaining a speech interval by comparing the speech power passed through the G C circuit with a threshold set based on the noise power during non-utterance, and a means for obtaining a speech interval when there is a silent interval of a certain length, In the speech recognition device, the threshold value is changed based on the noise power during the silent section after a fixed period of time after the end of the speech section within the word. This method is characterized by detecting the next audio section. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram showing an example of an electric circuit suitable for use in implementing the voice section detection method according to the present invention;
FIG. 2 is a signal waveform diagram for explaining the operating principle of the present invention. In FIG. 1, 1 is a microphone, 2 is an AGC (automatic gain control circuit), 3 is a power detection section, and 4 is a threshold variable section. ,
5 is an interval detection section. is silent section detection information due to consonants,
b represents threshold information.

To explain in detail with reference to FIG. 2, a speech section is detected using a threshold value B1 determined by the power of noise before speech is input, and in the illustrated example, "ST○
” is detected. If there is a silent section in the word,
Since the gain of the AGC circuit is smaller than that of the silent section, the power of the noise in the word is small. Generally, when the gain of the AGC circuit is increased to detect the end of a word, the time constant is approximately several seconds, so the gain is as shown in FIG. 2(b). Therefore, the gain of the AGC circuit does not change even during a silent section within a word. The threshold value B2 is set by focusing on the silent section within the word, and the following method is available as a method for setting the threshold value B2, for example.

(2) Set based on the noise power at time T, which is a certain period of time after the end of the speech section within the word.

■, from several tens of ms to 1. Set based on the average noise power of about OOms.

(2) The threshold value B2 is changed every fixed period of time, for example, 100 ms, until the next voice section is detected within the silent section.

FIG. 2(c) shows the voice section signal obtained as described above.

There are roughly two types of time between the end of a vocal section within a word and the start of the next vocal section. One type is a silent section of approximately 50m5 or less, which occurs before a consonant such as T'''IIK''.

The other one is due to consonants and is a silent section of 100 ms or more. Among these, in order to change the threshold only in silent sections due to consonant sounds, approximately 1. If a new voice section is detected within about OOms, the threshold value is not changed. Further, in the case of the end of a word, if a silent period continues for a certain period of time, for example, about 300 m5 to 400 ms, the end of the word is detected, but the gain is increased by the AGC circuit. In other words, it is necessary to change the threshold value periodically while the noise increases. The time for this change is determined by the time constant that increases the gain of the AGC circuit.

Effects As is clear from the explanation in section 2, according to the present invention, it is possible to stably detect a voice section without losing voice information even under noise.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an example of an electric circuit used to carry out the present invention, FIG. 2 is a signal waveform diagram for explaining the operating principle of the present invention, and FIG. FIG. 3 is a signal waveform diagram for explaining an example of a section detection method. DESCRIPTION OF SYMBOLS 1... Microphone, 2... AGC circuit, 3... Power detection part, 4... Threshold variable part, 5... Section detection part. 71

Claims (3)

[Claims] (1) An AGC circuit that removes differences in input voice power depending on the speaker, and a voice interval that compares the voice power that has passed through the AGC circuit with a threshold set based on the noise power during non-utterance. and, if there is a silent interval of a certain length, a means for generating a one-word signal that uses the silent interval as a word delimiter. A speech section detection method that detects the next speech section by changing a threshold value based on the noise power in a silent section that has elapsed for a certain period of time. (2), the noise power during a silent section after a certain period of time from the end of the vocal section within a word is calculated by several tens to 100 meters.
Claim (1) characterized in that the next voice section is detected by changing the threshold value based on the value averaged over s.
The speech interval detection method described in Section. (3) Detect the noise power during a silent section after a certain period of time has elapsed from the end of the speech section within a word, and change the threshold value sequentially based on the detected power each time. The voice interval detection method according to claim 1, wherein the voice interval detection method detects the voice interval that occurs at the end of the voice interval.