Reasearch Of Deep Learning In Audio Signal Processing

Signal Denoising has always been a tough task in signal processing.Most of the image processing methods before are based on SIFT,HOG and other manual pre-extraction features.Since the values of image data are always in range of 0-255 and based on matrix computation,it is more suitable for convolution operation.But audio signals are mostly one-dimensional,whose processing method is so different from the former.In the past,audio signal processing was often performed manually by means of third-party software.It was a tough task to deal and might do harm to the operator's hearing.Following the fast development of deep learning theories and hardware,complex model structures can be effectively used for the computation of signal processing.This project is both the final viva of Queen Mary University of London and Yamaha Music Co.Japan.As a member of Yamaha Music Co's summer international internship,aims to design a series of high-performance deep learning algorithms to realize the detection and reduction process in audio signal processing.In order to solve the noise detection and reduction problems in audio signal processing,aims to propose an algorithm module,then merge the module into a auto-noise reduction system.The whole module can be divided into two parts.The first is designed for the audio noise detection.I designed two algorithms for this,the CR-DNN model and Dense-DNN model.The second is generative model used to generate the clean audio file based on Generative Adversarial Networks which I called it Denoise-GAN.The performance measure of the former is accuracy,another is using both method of noise-evaluation like PESQ and Nash equilibrium in loss function.Not only using the traditional audio feature extraction methods,but also un-structured data ways like str-vectorization and word-vectorization during the generating of dataset provided by Yamaha.The audio dataset has been devided into training,testing and development dataset for different use.The experimental results show that deep learning can greatly improve the efficiency of noise detection and the Dense-DNN achieves more than 97% in accuracy.At the same time,the Denoise-GAN model can effectively enhance the clean audio portion of the original audio and greatly reduce noise interference.The model is currently used by Yamaha Music since Sep.2019.