Design of Speech Enhancement System Based on Microphone Array

Abstract

Beamforming is an important speech enhancement technology that can be employed to separate the speech signal from interfering noise. This technique has a decades-long history of application on microphone arrays, ranging from the basic Delay-Sum beamformer to the recently developed CGMM-MVDR beamformer based on iterative optimization algorithm.

This work evaluates the capability and stability of different beamforming methods and variously shaped microphone arrays in isolating human speech and enhancing speech intelligibility through simulation experiments. Furthermore, the issues and advantages of each method are discussed by examining the beampattern and the spectrum of the processed signal. In particular, this work discovered a deficiency in the Time-Frequency masking estimation of the CGMM-MVDR method, analyzed the causes of the issue, and proposed a post-processing method to process the mask, significantly improving the final result.

In addition, based on the results of the simulation experiments, this work designed and fabricated a prototype that includes a microphone array and data processing devices, presenting a viable solution for implementing the improved CGMM-MVDR beamforming algorithm on a portable device. An Android app was also designed to control the steering direction of the beamformer via gesture interaction, and simple tests were carried out in real-world scenarios.