Thung, You Jun
Date of Issue2016
School of Electrical and Electronic Engineering
Eavesdropping technology is popular not just for security surveillance but also for commercial applications such as for the use in meeting/conferencing scenarios. This project analyses a newly developed algorithm, the Direction-Informed Speech Enhancement (DISE) algorithm which uses the DOA of the desired signal to calculate the Hermitian angle between the signal and a direction vector. A mask is then applied to the received signal in the TF domain, which is a typical approach in UCBSS. The DISE algorithm performs substantially better over a fixed beamformer using the same physical setup. However, there were weaknesses in this algorithm which were the presence of artifacts (musical noise) in the processed signal and the degradation of performance with increased reverberation. Solutions to improve the masking method were proposed, based on the SSC measure used in UCBSS. The proposed implementations of the SSC measure proved to be an improvement to the algorithm performance in terms of speech clarity. Despite the trade-off between source isolation and signal distortion due to the masking, this new implementation helps to reduce this trade-off effect and allow the processed speech signal to be clearer without too much compromise on the SIR.
DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing
Final Year Project (FYP)
Nanyang Technological University