Speech dereverberation method based on depth characteristics of generative adversarial network

Abstract

The invention discloses a speech dereverberation method based on depth characteristics of a generative adversarial network, the method mainly comprises the following steps of: carrying out signal preprocessing on speech through weighted prediction error (WPE); carrying out characteristic extraction of data: carrying out MFCC characteristic extraction on speech data and bottleneck characteristic BNF extraction aiming at phonemes without reverberation speech; then carrying out signal preprocessing on speech through WPE: constructing the generative adversarial network with depth characteristics containing the MFCC characteristics of reverberation speech mapped to clean speech through the generative adversarial network; finally carrying out forced alignment by using a traditional GMM-HMM through a Kaldi toolbox, and then carrying out acoustic model training and decoding by using a depth neural network. The speech dereverberation method combines a method of signal processing with a deep learning framework based on the generative adversarial network to enable a system to combine the respective advantages of both to produce a better speech dereverberation effect.

Description

technical field [0001] The invention relates to the field of speech signal processing, and in particular aims at the problem of recognition performance degradation caused by environmental reverberation in far-field speech recognition, and proposes a speech reverberation method based on deep features of generative adversarial networks. Background technique [0002] In recent years, emerging industries such as smart home, dialogue robots, and smart audio have flourished, which has brought about great changes in people's lifestyles and the way people interact with machines. Voice interaction, as a new way of interaction, has gained popularity in these emerging fields a wide range of applications. With the application of deep learning in speech recognition, the recognition performance has been greatly improved, the recognition rate has exceeded 95%, and the recognition effect has basically reached the level of human hearing. However, the above are limited to near-field conditio...

AI Technical Summary

Problems solved by technology

Although the current neural network method can establish a good nonlinear mapping, it is difficult to achieve the desired effect only by using a fully connected neural network. Secondly, the use of the most basic feature mapping method cannot learn speech very well. Semantic information, constructing a good network structure and studying the deep semantic information of speech will produce a good recognition performance improvement for speech recognition, which has practical significance for speech recognition in complex scenes

Images