Title:  Some New Understanding of Phase Information and its Application to Speech Processing

Committee: 

Dr. Lee, Advisor  

Dr. Juang, Chair

Dr. Clements

Abstract:

The objective of the proposed research is to exploring new understanding and possible means to retrieve phase information in speech signal based on spectral magnitude and partial or corrupted phase, and to study practical methods to recovery phase or reconstruct signal with modified magnitude. With the fast growing of deep neural network models, more and more tasks have been boosted when move on to deep models. Among them speech signal processing is one of the most province application. However, most speech processing tasks, e.g. speech enhancement, speech bandwidth extension, dereverberataion, and etc., will only focus on improving the estimation of spectral magnitude and model coefficients, and hardly touch the part of phase. Besides, end-to-end deep models become hot topic in recent years, and one key point of end-to-end models in speech society is to take time-domain signal sequences as the input. However, most time-domain convolutional deep models and complex spectral domain feedforward models are not working well. On the other hand, there are evidence showing the effort of using feature on phase, e.g. in speaker identification and automatic speech recognition. And thus proposed research is trying to fill the gap between spectra feature and time-domain signal by introducing some new findings and understanding of the impact of phase, and to leverage the success of magnitude based models by studying practical method of phase recovery. In the preliminary research, the effect of phase and its mechanism was studied and some constraints for phase recovery was explored, and they demonstrate the potential of phase and its possible application to speech processing, which would be further studied.