Speech enhancement algorithms based on Wiener filter and compressive sensing

Abstract

Due to the advanced technologies, speech enhancement has become one of the prominent driving force in communication. Currently, there is a strong need for innovative single-channel speech enhancement algorithms that perform well at various types of noise levels. In this thesis, the compressive sensing technique was examined and evaluated for its suitability to be incorporated in speech enhancement algorithms due to its ability to recover signal from far fewer samples. Two novel speech enhancement algorithms have been proposed. The first algorithm was developed based on the modification of Wiener filter approach and compressive sensing. While the second algorithm added post-processing method using Gammatone filter to further improve the previous algorithm. Objective assessment test using Perceptual Evaluation of Speech Quality (PESQ), i.e. ITU-T P.862 standard, demonstrates that the first algorithm achieves around 15.19% improvement which outperforms other 16 traditional algorithms across 15 speech signals, 8 types of noise, and 0 to 15 dB SNR levels from NOIZEUS database. Moreover, the second algorithm further enhances the results to 16.38% improvement on average. Hence, the proposed speech enhancement algorithms based on compressive sensing have good potential across many types and intensities of environmental noise.