Speech separation method and system based on Ultra Gaussian prior speech model and deep learning and storage medium

Abstract

The invention provides a speech separation method and system based on Ultra Gaussian prior speech model and deep learning and a storage medium. The speech separation method comprises the steps of: utilizing a pure speech power spectrum density estimated value and a noise power spectrum density estimated value so as to obtain a prior signal-to-noise ratio in a gain function; bringing the prior signal-to-noise ratio into the gain function to obtain a value of the gain function; multiplying the value of the gain function by noisy speech spectrum to obtain an estimated value of a pure speech amplitude spectrum; and by utilizing an overlapping-adding technology, obtaining a recovered speech signal. The speech separation method and system and the storage medium have the beneficial effects that by combining a conventional statistic model with a deep learning technology, not only can a non-stationary noise signal be effectively inhibited, but also a problem of weak generalization ability caused by high dependence of the deep learning technology on training data is solved. Combination of the conventional statistic model and the deep learning technology enables enhancement performance of themethod to show very robust in various noise environments and signal-to-noise ratio cases.

Description

technical field [0001] The invention relates to the technical field of speech processing, in particular to a speech separation method, system and storage medium based on a super-Gaussian prior speech model and deep learning. Background technique [0002] Since speech signals are often polluted by interfering noise from the surroundings, applications such as automatic speech recognition, human-machine dialogue, and hearing aids have encountered great challenges. The performance of existing traditional speech enhancement technology is greatly reduced for non-stationary noise and low signal-to-noise ratio. Although the recently emerging deep learning-based speech enhancement technology can suppress non-stationary noise well, the performance of such algorithms is highly dependent on training data, and it will perform poorly for unlearned or trained data. Contents of the invention [0003] The invention provides a voice separation method based on a super-Gaussian prior voice m...

AI Technical Summary

Problems solved by technology

[0002] Applications such as automatic speech recognition, human-machine dialogue, hearing aids, etc. have encountered great challenges because speech signals are often polluted by interfering noise from the surroundings

The performance of existing traditional speech enhancement techniques is greatly reduced for non-stationary noise and low signal-to-noise ratio

Although the recently emerging speech enhancement technology based on deep learning can suppress non-stationary noise well, the performance of such algorithms is highly dependent on training data, and it will perform poorly for unlearned or trained data.

Images