High-perform speech enhancement method based on deep learning

Abstract

The invention relates to a high-performance speech enhancement method based on deep learning, which comprises the steps of preprocessing an audio PCM coding signal; performing framing and windowing onthe audio PCM coding signal, and dividing an original data set into a training set and a test set according to a set proportion at the same time; determining the number of DBN hidden layer nodes by using a golden section method; pre-training a weight and an offset parameter of the DBN by using the training set; perform fine tuning on the weight and the offset parameter of the DBN by using the training set; extracting DBN features, namely, respectively extracting features of the DBN training set and features of the DBN test set by using the DBN weight and offset parameter obtained by the finetuning; training a supervised learning speech separation system by using the extracted features of the DBN training set; enabling the extracted features of the DBN test set to serve as the input of the trained supervised learning speech separation system, wherein the output of the supervised learning speech separation system is an estimated target label, and the final enhanced speech can be obtained through speech waveform synthesis. The high-performance speech enhancement method has been greatly improved in speech evaluation index.

Description

technical field [0001] The invention relates to a voice enhancement method. In particular, it relates to a high-performance speech enhancement method based on deep learning. Background technique [0002] 1. The working principle of the supervised learning speech separation baseline system [0003] Computational auditory scene analysis in a supervised learning speech separation system is a typical speech separation method, which is based on the perceptual principles of auditory scene analysis, usually with ideal ratio masks as training targets for noise suppression. The ideal ratio mask is a time-frequency mask constructed from premixed speech and noise, defined as follows, [0004] [0005] Among them, S 2 (t, f), N 2 (t, f) represent the energy of speech and noise in the time-frequency unit, respectively, and the time-frequency unit is the expression after the signal passes through the γ-tone filter bank and the sub-band signal is divided into frames. β is an adjust...

AI Technical Summary

Problems solved by technology

[0012] Each RBM is pre-trained using the above contrastive hashing algorithm to obtain the initial weight of the DBN. However, the initial weight obtained by pre-training may only consider the local optimum of the cost function, ignoring the global nature. In order to make up for this shortcoming, it is necessary to Perform overall network parameter fine-tuning

Images