Deep Learning Based Multi-channel Speech Enhancement Methods

Speech enhancement techniques can suppress noise,improve the intelligibility and perception of speech,which is an important part of speech interaction.In recent years,many devices are equipped with multiple microphones to collect sound signals,which makes multi-channel speech enhancement algorithms more important.The traditional multi-channel speech enhancement methods are mainly based on beamforming algorithms.But it is difficult to estimate the parameters accurately and the performance decreases significantly in practical scenarios.Recently,deep learning based multi-channel speech enhancement methods have achieved significant results.In this thesis,we use the deep learning methods to study multi-channel speech enhancement problems and proposed the following improvements in view of their disadvantages:(1)Neural convolution-and-sum beamformer,we transform the traditional beamforming methods into a filter estimation problem.The performance of the speech enhancement task is improved by combining deep learning methods with traditional beamforming frameworks.And different neural beamformers can be implemented by using deep neural networks to estimate filters;(2)End-to-end multi-channel speech enhancement neural network,mapping based methods can make better use of information between channels by adding 1×1 convolution.It also alleviates the white noise problem which affects speech quality when STOI is used as loss function.Experiments on CHi ME3 dataset demonstrate that our proposed two improved methods can achieve better results in speech intelligibility and perception than traditional methods and similar deep learning based methods.