Blind speech separation : algorithm improvement and implementation using raspberry pi with uma-8-sp mic array testbed

Abstract

In a real-world environment, microphones record not only the target speech signal but also other available sources, the room acoustic effects, and background noise. Hence, extracting target speech from noisy convolutive mixtures is highly desirable for many applictions. This work aims to address the convolutive blind source separation of speech signals. First, we studied and compared three frequency-domain blind speech separation algorithms: IVA, Fast IVA, and ILRMA. Then, we worked on improving the performances of these algorithms using two different post-processings: speech denoising and SIMO equalization. The results demonstrate a significant improvement in performance. Finally, the selected separation scheme was implemented on an embedded system and tested on real-world signals