End-to-end Blind Speech Enhancement with Atrous Causal Convolutional Generative Adversarial Networks

Abstract

The invention relates to the fields of artificial intelligence and medical rehabilitation equipment. In order to propose an end-to-end bone conduction voice enhancement method and solve the problems of lack of high-frequency components of bone conduction voice, poor hearing and perception, and communication under strong noise background, the invention discloses An end-to-end bone conduction speech blind enhancement method based on hole causal convolution generation confrontation network. The method takes bone conduction original audio sampling points as input data, pure air conduction original audio as training output target, and trains bone conduction speech input well. The atrous causal convolution generates an adversarial enhancement network. The atrous causal convolution generates an adversarial network including a generator and a discriminator. The generator uses atrous causal convolution to output enhanced samples; the discriminator inputs the original audio data and generates The enhanced speech samples are generated by the detector, and the deep nonlinear features are extracted by the convolution layer in the discriminator, so as to judge the deep similarity of the samples. The invention is mainly used in the design and manufacture of bone conduction speech enhancement equipment.

Description

technical field [0001] The invention relates to the field of artificial intelligence, in particular to a training method and system for an end-to-end speech enhancement model. Specifically, it relates to an end-to-end bone-conducted speech-blind enhancement method for generative adversarial networks with dilated causal convolutions. Background technique [0002] The difference between bone conduction microphone (Bone Conducted Microphone, BCM) and traditional air conduction microphone (Air Conducted Microphone, ACM) is that the sound collected by BCM is not transmitted through the air, but through the highly sensitive vibration sensor to collect the human bone or tissue vibrations, which are then converted into audio signals. Its advantage is that it shields the noise from the sound source and prevents the transmission of environmental noise at both ends of the communication system from the source. Therefore, even in a strong noise environment, useful signals can be clearly...

AI Technical Summary

Problems solved by technology

[0007] In order to overcome the deficiencies in the prior art, the present invention aims to propose an end-to-end bone conduction speech enhancement algorithm based on the hole causal convolution generation confrontation network, and the proposed system uses end-to-end (ie waveform input and waveform output) Speech enhancement is carried out by means of training, the best network model parameters are obtained through training, and then the trained model is used to enhance bone conduction speech, so as to solve the problems of lack of high-frequency components of bone conduction speech, poor auditory perception and communication under strong noise background

Images