Multi-channel Speech Separation System Based on Gated Recursive Fusion of Deep Embedded Features

Abstract

The invention discloses a multi-channel speech separation system based on gated recursive fusion of deep embedded features, including a gated recursive fusion module, a deep embedded feature extraction module, a speech separation module, a discriminative training module and a joint training module, and the gated recursive The fusion module deeply fuses the spatial information and the amplitude spectrum information, and outputs the gated recursive fusion feature; the deep embedded feature extraction module, through the deeply embedded feature loss objective function, extracts more distinguishing features from the gated recursive fusion feature The deep embedded features; the voice separation module separates the deep embedded features to obtain each source and target speech signal; the discriminative training module obtains the discriminative loss objective function through the differentiated source and target speech signals; the joint training module passes Joint Training of Discriminative Loss Objective Function and Deeply Embedded Feature Loss Objective Function.

Description

technical field [0001] The invention relates to the technical field of signal processing, in particular to a multi-channel speech separation system based on gated recursive fusion of deep embedded features. Background technique [0002] Speech is one of the main means for human to exchange information, and speech separation has always occupied an important position in speech signal processing. Speech separation is also known as the cocktail conference problem, and its goal is to separate each target source speech signal from a speech signal containing multiple mixed speakers. When a piece of speech contains multiple speakers at the same time, it will seriously affect the performance of systems such as speech recognition, speaker recognition and hearing aids, so speech separation technology is particularly important. In the development process of speech separation technology, many speech separation methods based on deep learning have achieved good results, such as deep clust...

AI Technical Summary

Problems solved by technology

First, MDC only stitches the spatial information as an auxiliary feature to the magnitude spectrum feature, which makes it difficult to learn the mutual information between the space and the magnitude spectrum, and the distribution of IPDs and the magnitude spectrum feature is different. It is also not conducive to the learning and optimization of the network

Second, the training objective function of MDC is defined on the deep embedding vector, not on the real separation target. These deep embedding vectors cannot perfectly represent the target speech, so it will damage the performance of speech separation.

Images