Premium
On enhancing model‐based expectation maximization source separation in dynamic reverberant conditions using automatic Clifton effect
Author(s) -
Gul Sania,
Khan Muhammad Salman,
Shah Syed Waqar,
Lloret Jaime
Publication year - 2020
Publication title -
international journal of communication systems
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.344
H-Index - 49
eISSN - 1099-1131
pISSN - 1074-5351
DOI - 10.1002/dac.4210
Subject(s) - computer science , pesq , reverberation , distortion (music) , source separation , speech recognition , separation (statistics) , maximization , algorithm , similarity (geometry) , signal (programming language) , perception , artificial intelligence , acoustics , speech enhancement , mathematical optimization , image (mathematics) , telecommunications , machine learning , mathematics , amplifier , physics , bandwidth (computing) , neuroscience , noise reduction , biology , programming language
Summary Source separation algorithms based on spatial cues generally face two major problems. The first one is their general performance degradation in reverberant environments and the second is their inability to differentiate closely located sources due to similarity of their spatial cues. The latter problem gets amplified in highly reverberant environments as reverberations have a distorting effect on spatial cues. In this paper, we have proposed a separation algorithm, in which inside an enclosure, the distortions due to reverberations in a spatial cue based source separation algorithm namely model‐based expectation‐maximization source separation and localization (MESSL) are minimized by using the Precedence effect. The Precedence effect acts as a gatekeeper which restricts the reverberations entering the separation system resulting in its improved separation performance. And this effect is automatically transformed into the Clifton effect to deal with the dynamic acoustic conditions. Our proposed algorithm has shown improved performance over MESSL in all kinds of reverberant conditions including closely located sources. On average, 22.55% improvement in SDR (signal to distortion ratio) and 15% in PESQ (perceptual evaluation of speech quality) is observed by using the Clifton effect to tackle dynamic reverberant conditions.