An ERP Study to Evaluate the Quality of Speech Processed by Wiener Filtering

Abstract

Speech enhancement algorithms (SEAs) play a crucial role in speech signal processing. Wiener filtering (WF) as one of the most popular SEAs introduces undesired speech distortion to the enhanced signals. It is important to objectively examine the perception of the enhanced speech in different conditions (i.e., signal-to-noise ratio (SNR) level). The purpose of this work was to investigate whether the mismatch negativity (MMN) response based on the event-related potential (ERP) could objectively evaluate the quality of speech processed by WF. There were two deviant stimulus types: the noisy stimuli corrupted by a steady-state noise at SNR of -5 and +5 dB and their corresponding enhanced stimuli processed by WF. Experimental results showed that the MMN amplitude elicited by the noisy stimuli at -5 dB SNR was larger than that at +5 dB SNR, accounting for the effect of SNR level on the quality of speech. The results also demonstrated that the enhanced stimuli at -5 dB SNR evoked a shorter MMN latency and a larger MMN amplitude than that at +5 dB SNR, reflecting the effect of SNR level on the quality of WF-processed speech. The MMN may be potentially used as an objective measure for evaluating the quality of speech processed by WF.