Multi-channel speaker-independent voice separation method based on deep clustering

Abstract

The invention provides a multi-channel speaker-independent voice separation method based on deep clustering. The method comprises the following steps: firstly, carrying out short-time Fourier transform on a voice signal to extract amplitude spectrum characteristics of the voice signal, then calculating cosine values of phase differences between different channels to serve as spatial characteristics, and combining the two characteristics to serve as input characteristics for training a deep clustering network; then, training a bidirectional long-short-term memory network, and obtaining estimated masks of different speakers by utilizing the network; and finally, calculating the coefficient of the MVDR beamformer by using the spatial covariance matrix, and multiplying the mixed voice by the obtained beamformer coefficient to obtain separated speaker voice signals. According to the method, the spatial information of the voice signals is better utilized, the high-quality mask is estimated by using the deep clustering network, the separation processing of the mixed voice signals of a plurality of speakers in the reverberation environment can be realized, and the method has better voice separation performance.

Description

technical field [0001] The invention belongs to the technical field of speech signal processing, and in particular relates to a multi-channel and speaker-independent speech separation method based on deep clustering. Background technique [0002] Speech separation is a relatively basic task in the signal field, and it is a special case of sound source separation. Its goal is to separate the target speech from the background noise. Speech separation has many applications, including hearing prostheses, communications, automatic speech processing, and speaker recognition, among others. For the human hearing system, even in a cocktail scene, we can easily hear a person's speech from the voices of other people and the noisy background noise around them. Therefore, the speech separation problem is often referred to as the "cocktail party problem". But the reason why humans can easily separate speech is that before the sound signal is transmitted to the human auditory center, it ...

AI Technical Summary

Problems solved by technology

However, the robustness of the former is poor, and the latter has large nonlinear distortion, so they are not particularly ideal, and there is still room for improvement.

Images