Single channel mixed speech time domain separation method based on Convolutional Neural Network

Abstract

The invention relates to a single channel mixed speech time domain separation method which is the single channel mixed speech time domain separation method based on a Convolutional Neural Network. The method comprises the following steps of 1, constructing a speech data set for training, 2, preprocessing speech data, 3, obtaining hybrid speech data, 4, constructing a neural network structure, 5, using data organized to train a neural network in a monitoring mode, and 6, using the trained neural network to carry out a separation test. According to the single channel mixed speech time domain separation method based on the Convolutional Neural Network, time domain speech signals serve as input and output of the Convolutional Neural Network, single channel hybrid speeches are separated, and therefore estimation for two source signals is obtained. The method does not need to deal with the problem of phase retrieval, and the separation quality of a single channel speech is improved.

Description

Technical field [0001] The invention relates to a time-domain separation method for single-channel mixed speech, and more specifically, to a time-domain separation method for single-channel mixed speech based on a convolutional neural network. Background technique [0002] Single-channel blind source separation (Monaural Blind Source Separation, MBSS) is an important technology in the field of speech processing. It can obtain estimates of two-channel source signals when only a single-channel mixed speech signal is obtained. Single-channel speech separation technology has important application value in speech recognition, speech enhancement, speech identification and other fields. [0003] Typical single-channel speech separation includes methods based on non-negative matrix factorization (NMF) and neural networks (Neural Network). Since the single-channel mixed speech contains less information, it is difficult to achieve satisfactory separation results based on non-negative matrix...

AI Technical Summary

Problems solved by technology

However, the current neural network-based methods generally use a fully connected network (Full Connected Neural Network, FCNN) or a recurrent neural network (Recurrent Neural Network, RNN), and usually need to extract the amplitude spectrum features of the speech signal, which has not been well utilized. To the powerful feature expression ability of the convolutional neural network itself; at the same time, due to the use of the amplitude spectrum feature, it is faced with the difficult phase recovery problem when restoring the source signal

Therefore, the traditional neural network-based separation method has mutual interference between the separated two source signal estimates, and the separation quality needs to be improved.

Images