Evaluation Of The Quality Of The Noise-Suppressed Speech Using EEG Signal

The existence of various noise interferences seriously affects speech quality,and then affects the effective communication between people.Noise-suppression technology can recover clean speech as much as possible from a noisy speech signal.An important method to evaluate its performance is to evaluate the quality of the noise-suppressed speech signals.Cochlear implant(CI)is the medical intervention device that can help profoundly deafened patients to restore and improve their hearing,but speech perception in background noise remains a challenge in CI users.Considering the different development of auditory mechanism between CI users and normal-hearing(NH)persons,it is crucial to establish speech noise-suppression algorithm that meets the auditory perception of CI users,and after all,it is significant to evaluate their perception of speech quality.In order to study the mechanism of human auditory information perception and cognitive processing,electroencephalogram(EEG)has been proposed as a new objective and direct method to evaluate the auditory perception of speech quality.Event-related potential(ERP)in EEG analysis is a brain response with phase-locked time specific events.ERPs provide a valuable means for studying cerebral perception and auditory information recognition.Mismatch negativity(MMN)is an ERP component that reflects the brain's pre-attention processing of information and is an automatic process of detection and processing.This study explored the basic characteristics of MMN in NH persons and CI users,and then analyzed their ability to distinguish speech quality.Firstly,this dissertation introduced the basic theories of EEG,the principles of CI,the influence of noise on CI users and speech noise-suppression algorithms,and then discussed the methods of evaluating speech perception of CI users by EEG.In this study,15 NH subjects and 5 CI users were tested for MMN.In the odd-ball paradigm,the clean vowel /a/ stimulus was used as the standard stimulus,and the noisy vowels at signal-to-noise ratio(SNR)levels of-5dB and +5dB and the noise-suppressed vowels processed by Wiener filtering were used as the deviant stimuli.The differences of amplitude,latency and brain topographic map of MMN were analyzed,and the effects of SNR level and noise-suppression processing on MMN were explored.The results showed that the amplitude of MMN increased with the decrease of SNR in NH persons and CI users,accounting for the effect of large quality difference between the speech at low SNR and the clean speech.These findings suggested that MMN(particularly its amplitude)could be potentially used as an objective tool to evaluate the quality of speech.The brain activation areas of responding to noisy speech and noise-suppressed speech in NH persons were both concentrated in the frontal lobe and parietal lobe,and the brain topographic map showed symmetrical distribution.However,the brain activation area of processing the noisy speech in CI users was on the left side of the parietal lobe,and the brain activation area of processing the noise-suppressed speech was on the right side of the frontal lobe.The brain topographic map showed the lateralization effect which reflected the different automatic processing mechanism of the brain between the noisy speech and the noise-suppressed speech,providing the neurophysiological basis for improving the speech noise-suppression technology in line with the auditory perception of CI users in the future.