Voice signal detection method of digital audio system

Abstract

The invention discloses a voice signal detection method of a digital audio system. The method comprises: a sample mean absolute value is obtained from each frame of voice signal and an average absolute deviation value of all voice signals is calculated based on a plurality of sample mean absolute values of multiple frames of voice signals; if the average absolute deviation value is less than a noise threshold value all the time within set noise detection time, all detected voice signals are determines as noises; and if the average absolute deviation value is larger than the noise threshold value all the time within set voice detection time, the detected voice signals are determined as voice signals; and under other circumstances, the detected voice signals are determined as ones in intermediate states. According to the invention, the average absolute deviation value is used as a detection decision metric and obvious differences between the voice signals and the noises in terms of energy and stationary performances are used. Therefore, the computation complexity is reduced; the requirement on hardware during realization is reduced; and the good detection performance can be guaranteed.

Description

technical field [0001] The invention relates to digital audio system signal detection, in particular to a voice signal detection method. Background technique [0002] In the digital audio system, the processing of the speech signal is based on the activation detection, and it is necessary to distinguish the speech signal from the noise. [0003] The current detection methods mainly use the statistical characteristics of the speech signal, such as amplitude, energy, zero-crossing rate, quasi-periodicity, frequency characteristics, correlation, etc., and judge according to the maximum likelihood criterion. These methods all extract the characteristic parameters that can distinguish speech from noise or transform them to obtain obvious difference results, so as to find out the dividing point between the two. [0004] The energy of the speech signal is always greater than the background noise under high signal-to-noise ratio, and the energy detection method has better detection...

AI Technical Summary

Problems solved by technology

Due to the influence of mixed speech, especially the zero-crossing rate of unvoiced sounds is comparable to noise, the correct detection probability of the zero-crossing rate method is low in some speech environments

[0006] In speech signal detection, methods based on amplitude, energy, and zero-crossing rate are simple to implement, but their detection performance is not good. Although methods based on quasi-periodicity, correlation characteristics, and frequency characteristics can achieve good detection performance, the amount of calculation is very large. Large, high hardware requirements

Existing detection methods are often based on binary judgment criteria, and only one threshold variable is used to judge and distinguish. The setting of the single judgment threshold is the most important. The judgment result is either a voice signal or noise, and the probability of false alarms and false alarms is high.

Images