Speech Bandwidth Extension Based On Neural Network

Compared with the wideband speech signal,the narrowband speech signal losses the information of the high-frequency part.The purpose of the speech bandwidth extension task(BWE)is to reconstruct the missing information of the narrowband speech.In recent years,the BWE has been greatly developed with the introduction of the neural network.In the supervised learning method,use the characteristics of the narrowband speech as the input data of the neural network,and obtain the estimated wideband speech at the output of the neural network.The neural network established a mapping relationship between the narrowband speech and the wideband speech,which further reduces the complexity of the BWE.Based on the neural network mapping,this thesis investigates the BWE from two aspects:1.Clean speech bandwidth extension.Clean speech bandwidth extension is the focus of the BWE,and this thesis used a composite neural network that consists of a convolutional neural network and long short-term memory to implement the BWE task.The composite neural network combines the advantages of the convolutional neural network and long short-term memory,which is helpful for processing long sequence speech data with strong relevance.This thesis proposed a data processing method based on the transformation relationship between the wideband speech,the narrowband speech,and the interpolation speech,called sampling-point replacement.This method can retain the information of the narrowband speech in the predicted wideband speech without affecting the BWE effect.2.Noisy speech bandwidth extension.The speech in daily life is usually accompanied by noise,the main research object of BWE is clean speech and there is less research on noisy speech.This thesis took the noisy narrowband speech as the research object and used two models of multi-task learning and single-task learning for data modeling to realize the BWE task of noisy narrowband speech.