Voice extraction method under high background noise

Abstract

The invention discloses a voice extraction method under high background noise. At first, original voice signals containing the high background noise are discretely sampled in advance and quantized to obtain data frames, a wavelet neural network based on a Morlet wavelet function is constructed for the data frames, a particle swarm fitness function is constructed for parameters of the wavelet neural network, then the optimal parameter is obtained through a particle swarm algorithm, the data frames are input to the wavelet neural network to be filtered, and therefore the noise is removed, and voice signals are extracted. The parameters of the wavelet neural network are obtained by means of the particle swarm algorithm. Therefore, fitness for different environmental noise characteristics can be improved through the method.

Description

technical field [0001] The invention belongs to the technical field of speech extraction, and more specifically relates to a speech extraction method under strong background noise. Background technique [0002] In some occasions with strong background noise, such as beside the highway, in roaring vehicles, among noisy crowds, etc., how to eliminate the strong background noise and effectively extract the user's voice is the speech recognition device in these special occasions. One of the main problems. [0003] At present, the noise reduction technologies mainly used in various voice communication devices include: [0004] 1) Model-based noise reduction methods, typically such as the speech model and Noise model, which filters and compensates the received audio signal by setting adaptation rules. This method has a good denoising effect in the case of a high degree of adaptation, but requires more prior knowledge to pre-build user speech and noise models. [0005] 2) Activ...

AI Technical Summary

Problems solved by technology

A typical example is "Northwestern Polytechnical University. A speech noise reduction method based on wavelet transform and variable step size least mean square algorithm. Chinese invention patent, CN101894561A, 2010.11.24" disclosed speech noise reduction method, which makes full use of wavelet Transformation has advantages in time-frequency local analysis, and introduces a step size dynamic algorithm to control the convergence speed, which has a faster response time and out-of-schedule, but the determination of the step size factor still lacks an effective method, which limits its application to different noise environments. adaptability

Images