Multichannel speech enhancement for automatic speech recognition: a literature review

Area of Science:

• Signal Processing
• Artificial Intelligence
• Acoustics

Background:

• Multichannel speech enhancement (MCSE) is vital for robust automatic speech recognition (ASR).
• Existing reviews lack a systematic analysis of MCSE specifically for ASR applications.
• Rapid advancements in MCSE methods, models, and datasets necessitate a comprehensive overview.

Purpose of the Study:

• To systematically review MCSE approaches for ASR systems.
• To analyze the performance of MCSE and ASR across various techniques, models, and noise conditions.
• To discuss current challenges, limitations, and future research directions in MCSE for ASR.

Main Methods:

• Systematic literature review using keyword searches across major electronic databases (Google Scholar, IEEE Xplore, etc.).
• Inclusion and exclusion criteria applied to 240 initial articles, resulting in 40 final experimental articles (23 journals, 17 conferences).
• Backward snowballing and quality assessment were used to finalize the article selection.

Main Results:

• An increasing trend in MCSE for ASR research was observed, with Word Error Rate (WER), Perceptual Evaluation of Speech Quality (PESQ), and Short-Time Objective Intelligence (STOI) as common performance metrics.
• A significant challenge identified is the lack of generality and comparability across MCSE studies, hindering unified solutions for speech recognition noise.
• Key MCSE methods that enhance ASR performance across diverse models, techniques, noise types, and environments were identified.

Conclusions:

• This review provides a comprehensive examination of MCSE and ASR techniques.
• Identified MCSE methods offer potential for improving ASR performance under various conditions.
• Future research should address the generality and comparability issues to develop more robust and unified MCSE solutions for ASR.