Summary: | 碩士 === 國立交通大學 === 電信工程研究所 === 107 === Speech enhancement has always been a very important issue in the pre-processing of speech signals. In the past few years, there have been many ways to reduce noise and eliminate reverberation. Unfortunately, traditional methods require some assumptions about mathematical models, but in real situations they are not necessarily correct. Machine learning solves this problem, therefore this paper simulates a variety of different environments, and enables neural networks to learn these information. Firstly, we estimate the DOA and use MVDR to enhance the speech in the target direction. And then combine enhanced signal and original signal as the input to LSTM to achieve the masking method. When the neural network is trained, the objective function is calculated in both frequency domain and time domain. For machine, it is hoped to improve the SNR. From the perspective of human hearing, the speech comprehension and speech quality are the evaluation criteria. Experiments have confirmed that the system can operate effectively in the environment of noise and reverberation, and the proposed method can indeed improve system performance.
|