A comparison of Improved Meta-Heuristic Optimisation Based Speech-Enhancement Systems

Abstract

Speech enhancement has become an area of interest to researchers in the field of machine learning as a result of the development of digital signal processing applications. The goal of speech enhancement is to increase the quality of these digital speech devices and update them to handle all sorts of noises. In the case of dual-channel speech enhancement systems, an Adaptive Noise Cancellation (ANC) system contains an adaptive filter and an adaptation algorithm (optimisation procedure) for adjusting their parameters. This thesis investigates the use of meta-heuristic methods to propose a novel speech enhancement systems. This results in exploiting different meta-heuristic optimisation techniques, such as Particle Swarm Optimisation (PSO), Gaussian Par- ticle Swarm Optimisation (GPSO), Accelerated Particle Swarm Optimi- sation (APSO), Bat Optimisation (BA), Gravitational Search Algorithms (GSA). This thesis presents the formulation of an ANC system based on Butterworth, and Elliptic filters, in the form of an optimisation task. PSO, GPSO, APSO, GSA,and BA are used to find the optimal filter coefficients, that optimise the perceptual evaluation of speech quality (PESQ), sig- nal distortion (C sig), overall signal quality (C ovrl), and Log-Likelihood Ratio (LLR), for the noise-free audio signal and the filtered signal. This made it possible to build a system capable of enhancing speech, where we employed different sentences from the NOIZEUS data-set and ARABIC speech corpus under various Signal to Noise Ratio (SNR) values. Objec-

tive and subjective evaluation tests, were conducted on the meta-heuristic speech enhancement systems and showed encouraging results. The results of the proposed speech enhancement systems revealed that PSO generally outperformed APSO and GPSO at all levels of SNR. The optimised El- liptic filter by BA showed improved scores compared to the fixed filter, and the audio-only Wiener filter at all SNR levels. The results confirmed that meta-heuristic based acoustic noise cancella- tion models considered are capable of high performance.