Voice detection method under noise condition

Abstract

The invention provides a voice detection method under noise condition, and belongs to the technical fields of digital signal processing, computer artificial intelligence and pattern recognition. The method comprises the following steps of: converting input signals to a frequency domain, and dividing into subbands; calculating a power spectrum of each subband to form a subband power spectrum; calculating mean-square deviation of the subband power spectrum of each frame, and comparing the mean-square deviation serving as a detection characteristic with an adaptive voice detection threshold to determine whether the current frame contains voice signals; and according to a detection result, adopting a certain endpoint determination strategy to determine an initial position and an ending position of a voice segment.

Description

technical field [0001] The invention relates to a digital signal processing, computer artificial intelligence and pattern recognition technology, in particular to a method for using a computer to detect voice in a signal. Background technique [0002] The accuracy of speech detection determines the performance of the entire speech processing system to a large extent. People have done a lot of research on speech detection and proposed many various methods. For example, based on short-term energy and short-time spectrum energy , short-term zero-crossing rate and other speech detection algorithms. But these characteristic parameters are sensitive to background noise and cannot describe the characteristics of speech well. Like short-term energy and short-term zero-crossing rate, they are not enough when the signal-to-noise ratio is low. To distinguish speech and background noise. Speech detection algorithms based on linear prediction coefficients, cepstral coefficients, and pitc...

AI Technical Summary

Problems solved by technology

For example, "A Threshold Adaptive Speech Detection System" disclosed in the patent document with application number 200310103263.7 can automatically update the threshold, but needs fuzzy clustering processing and Bayesian information processing, so a large amount of calculation is required. And without considering the characteristics of the voice itself, it is difficult to guarantee the performance under complex background noise and low signal-to-noise ratio

As disclosed in the "Speech Detection System for Noisy Environment" disclosed in the patent document with the application number 99104095.3, the sub-band division of the voice is carried out, but it is only divided into two parts, the high frequency band and the low frequency band, and the voice itself is not considered. Band characteristics

Images