A noise filtering framework in multi-channel speech enhancement system for environmental noises / Pavani Cherukuru

Pavani , Cherukuru (2023) A noise filtering framework in multi-channel speech enhancement system for environmental noises / Pavani Cherukuru. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (137Kb)
	PDF (Thesis PhD) Download (2402Kb)

Abstract

The speech enhancement system deals with noisy speech signals by reducing the background noises while preventing any alterations to the speech features. Speech enhancement algorithms are used in multiple channels applied in communication devices to enhance the quality of speech signals under noisy environments known as multi-channel speech enhancement system (MCSE). Micro Electro-Mechanical Systems (MEMS) microphones are used in MCSE systems in outdoor environments. There are many existing algorithms used to filter the noise in speech enhancement systems which are frequently used as a pre-processor to enhance speech quality. These algorithms were effective in the reduction of noisy signals and improved the quality of speech. However, they may have limited ability to perform well on low Signal-to-Noise Ratio (SNR) conditions. The existing MCSE systems can filter 0 to 60dB of SNR, which gives a 62.5% Word Recognition Rate (WRR) at 0dB (considered low SNR), and 83% WRR at 60dB (considered high SNR). However, it was tested only with white Gaussian noise but not with environmental noises, which is very crucial in speech communication devices. Thus, the existing MCSE did not consider all types of noises in a real-time environment. This research aims to propose a noise filtering framework using suitable algorithm(s) for multi-channel speech enhancement systems in handling various Signal-to-Noise ratio (SNR) of environmental noises. This research firstly analyzes the findings of the existing algorithms and components involved in the Speech Enhancement and MCSE systems in handling different types of noises. This is to identify suitable algorithms for proposing a noise filtering framework for environmental noises. Secondly, experiments were conducted on the existing MCSE as the benchmark systems to analyze the limitations of the existing algorithms in handling environmental noises. From the benchmark experiments, this research has identified that the MCSE’s recognition rate reported the highest WRR at 93.77% for high SNR (at 20dB) and 5.64% for low SNR (at -10dB) on an average of five types of different noises. This research has proposed a noise filtering framework that comprises the pre-processing and deep learning algorithms for MCSE in handling various SNRs of environmental noises. The performance of the developed noise filtering framework in handling various SNR of environmental noises shows a WRR of 70.55% at -10dB SNR and 75.44 % at 15dB SNR, while 5.82 % at -10dB and 88.8% at 15dB by the existing MCSE system. It has proven that the proposed pre-processing and deep learning algorithms performed well at low SNR’s for MCSE under noisy environments.

Actions (For repository staff only : Login required)