Speech endpoint detection method and system

Abstract

The invention discloses a speech endpoint detection method, comprising: obtaining multiple speech existence probabilities of multiple frequency points of the current audio frame obtained during the noise reduction process of the audio signal; determining the current audio frequency according to the multiple speech existence probabilities The existence probability of the voice signal of the frame; obtain the probability that the voice signal exists in each of the preceding L1 audio frames of the current audio frame, and determine the distance D1 of the audio frame with the largest probability value from the current audio frame; determine that there is a voice signal in the current audio frame Whether the average value of the probability and the sum of the probabilities of voice signals existing in each of the previous D1 audio frames is greater than the set threshold; if yes, it is determined that there are voice signals in the D1 audio frames behind the current audio frame. The present invention uses the voice existence probability of the audio frame during the audio signal noise reduction process to detect the voice endpoint, realizes simple statistical comparison using the signal processing result, greatly simplifies the calculation, and reduces the memory requirement.

Description

technical field [0001] The invention relates to the technical field of voice signal processing, in particular to a voice endpoint detection method and system. Background technique [0002] Voice activity detection (Voice Activity detection, VAD), also known as voice detection, is used in voice processing to detect the presence or absence of voice, thereby separating voice segments and non-voice segments in a signal. [0003] The current voice endpoint detection methods include: neural network method, double-threshold detection method, detection method based on autocorrelation maximum, and detection method based on wavelet transform. in, [0004] Neural network method: the features need to be designed manually, the implementation is more complicated, and the amount of calculation is relatively large. [0005] Double-threshold detection method: utilizes the short-term energy and short-term zero-crossing rate of speech, is suitable for scenes with high signal-to-noise ratio, ...

AI Technical Summary

Problems solved by technology

[0007] Detection method based on wavelet transform: slow detection speed and poor practicability

[0008] In the above prior art, there are also problems of large amount of calculation and high requirements for power consumption, processor performance and memory on embedded devices.

Images