An Integrated Speech Enhancement System Based on Multi-objective Heterogeneous Network

Abstract

The invention discloses an integrated speech enhancement system based on a multi-target heterogeneous network. The system includes a feature extraction module, a feature dimensionality reduction module, m heterogeneous networks, and n gating units, and uses m heterogeneous networks as an integrated The sub-model of the speech enhancement system, each heterogeneous network leads to multi-objective branches from the first network layer, and connects the first and last layers of the heterogeneous network in a symmetrical manner, which can effectively alleviate the parameter optimization conflict problem of multi-objective learning, and can Avoiding the loss of information in the original input in the deep network propagation can improve the diversity of the base model of the integrated speech enhancement system, thereby improving the quality and intelligibility of the enhanced speech. The feature dimensionality reduction module calculates the correlation information between the original input speech frames and concatenates it with the current input frame as the input of n gating units, which greatly reduces the input dimension of n gating units without losing the original Correlation information between frames in the input.

Description

technical field [0001] The invention relates to the technical field of speech enhancement, in particular to an integrated speech enhancement system based on a multi-target heterogeneous network. Background technique [0002] Speech is the most important and direct information carrier in people's daily communication. However, voice signals are often polluted by various noises in life, such as speaker noise in restaurants, machine noise in factories, construction site noise in construction sites, car noise on the road, noisy crowd noise, etc. These noises will affect our acquisition and understanding of useful speech, resulting in a decrease in the quality and intelligibility of speech. [0003] Speech enhancement technology refers to eliminating noise components from noisy speech, extracting and recovering clean speech components, so as to improve the listening quality and intelligibility of speech. The algorithms in it include traditional statistics-based augmentation tech...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to solve the technical defects of the existing multi-objective learning speech enhancement system and the integrated learning speech enhancement system, and provide an integrated speech enhancement system based on a multi-objective heterogeneous network, which can effectively alleviate the parameters of multi-objective learning Optimize the collision problem and avoid the loss of information in the original input in the deep network propagation

Images