Single-channel Speech Separation Algorithm Based on Deep Neural Network

Abstract

The invention provides a single-channel speech separation algorithm based on a deep neural network, which mainly includes the following steps: preprocessing the training speech samples and extracting their feature information; using a loss function to train the deep neural network to obtain a deep neural network Network model: Preprocess the speech sample to be tested, extract its feature information, and perform speech separation through the trained deep neural network model, and then obtain the separation result through speech reconstruction. The present invention uses the nonlinear relationship between input and output to train the deep neural network. Compared with the traditional separation method based on the single-output deep neural network, it fully excavates the joint relationship between the outputs, and the separation efficiency is high. Separating the two source speech signals effectively reduces the distortion rate of the speech and improves the intelligibility of the separated speech at the same time.

Description

Technical field [0001] The present invention relates to a single pass speech separation algorithm based on a depth neural network, belonging to the field of speech separation. Background technique [0002] Single Channel Speech Separation, SCSS is a process of recovering multiple speech from one-dimensional mixed speech. Single-channel speech separation techniques are widely used in voice enhancement, speech recognition pretreatment, hearing aid or smart home. In these areas, usually the sensor receives a mixed voice from a microphone, and the human ears can easily obtain useful information from this mixed speech, and for the computer, it is difficult to accurately get the desired voice. Therefore, it is very important to accurately and efficiently get a very important practical meaning. [0003] Deep Neural Network (DNN) has powerful data mining capabilities, in the field of speech separation, which is mainly used to simulate nonlinear relationships between input features and ou...

AI Technical Summary

Problems solved by technology

When using a single-output DNN to separate mixed voices, only one voice can be separated at a time. When using this method to separate multiple voices, it takes a long time; the traditional method based on a multi-output deep neural network can separate multiple voices at the same time. However, for this reason, the output mapped by the multi-output DNN is not as targeted as the single-output DNN, and the separation effect is worse than that of the single-output DNN.

Both of the above two deep neural networks need to be trained by a loss function. The basic loss function used by the traditional dual-output DNN is only used to map the relationship between input and output, but ignores the joint relationship between outputs, and this joint relation has a large impact on the final speech separation performance

Images