1. UKM Learning and Research Repository
2. Master Thesis / Tesis Master
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	Speech enhancement algorithms for Malay language based auditory training system
Authors:	Abrar Hussain (P69982)
Supervisor:	Kalaivani Chellappan, Dr.
Keywords:	Malay language Speech processing systems
Issue Date:	1-Mar-2018
Description:	Auditory Training System is a potential intervention to increase the perception of speech in noise and spatial processing ability of elderly individuals who have reduced hearing capacity due to ageing. In Malaysia, the main challenge in exercising auditory training is found to be the language limitation. There is no Malay language based Auditory Training System in the market currently. The purpose of this study is to design speech enhancement algorithms for real-time processing of the target speech in a home-based developed Auditory Training System by verifying the quality and intelligibility through a simulation platform. The developed simulation platform works under two modules: input speech verifying module and virtual acoustic control module. Input speech verifying module is employed with speech and noise separator; speech enhancer. Instantaneous mixing model and microphone array are used to design the input/mixing SNR level to prepare real-life acoustic environment for speech enhancer. Speech and noise separator guarantee accurate and high speed separation of speech and noise signals using multichannel source separation algorithms. The designed speech-noise separation index statistical validity was established based on projection pursuit and fast independent component analysis algorithms. Fast independent component analysis provided the highest quality of speech and noise separation index (r=1, p< 0.05). Fast independent component analysis also has the lowest computational demand of 4.2 ms which is suitable for real-time processing of the target word in an Auditory Training System. By employing single channel and multichannel noise reduction algorithms, speech enhancer ensures pre-recorded input speech (as training stimulus) quality and speech intelligibility while being tested it by mixing with the real-life noise signals of speech babble and competing speech. Single channel noise reduction was approached through two different algorithms namely restructured Ideal Binary Mask (IBM) and a combinational technique (Block-Thresholding algorithm and Phase Spectrum Compensation approach). Combinational technique is found to be superior in this approach by providing higher segmental signal-to-noise ratio of about 20 dB and providing less speech distortion while using most complex speech babble as noise/masker signal. Speech intelligibility is predicted using Short-Time Objective Intelligibility (STOI) and was found combinational technique produced good intelligibility of speech (STOI value of 0.5 and above) while extracted speech from speech babble. For multichannel noise reduction algorithms, time delay beamformer was found better compared to Frost beamformer in terms of moderate array gain of 6 (approximately) and speech distortion index close to 0. Virtual acoustic control module is employed with spatial audio simulator that controls spatial direction of speech signals and noise signals to verify spatial hearing ability. Finally, input speech verifying module is integrated with virtual acoustic control module. During the integration, enhanced speech from speech enhancer is used for spatial audio simulator. By performing the simulation analysis and informal listening test using spatial audio simulator it is confirmed that elderly listeners are very good at judging the target speech from 90 degree spatial position compared to 45 and 180 degree spatial positions.,Certification of Master's/Doctoral Thesis" is not available
Pages:	274
Call Number:	TK7882.S65H847 2018 3 tesis
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.