A Deep Learning Based Noise Robust Signal Processing Strategy For Cochlear Implants

A cochlear implant(CI)is a surgically auditory prosthesis that can partially restore hearing to listeners with severe to profound sensorineural hearing loss.Over the past decades,progresses in CI techniques have enabled CI recipients to enjoy near-to-normal speech perception in quiet environments.However,CI users still suffer from seriously degraded speech quality and intelligibility in noisy environments.To tackle such problems,conventional speech enhancement(SE)methods,as well as deep learning(DL)based ones,have been exploited and significantly improvements are obtained.However,these works are still far from satisfactory due to: 1)they are mostly exploited in the standard CI strategies as additional modules and their outputs are not guaranteed to match the CI patterns;2)most of the DL-based SE algorithms rely on well-trained models with large network size and high computational complexity,which restricts their practical CI implementation;3)the insufficient representation of temporal fine structures(TFS)in currently envelope-focused CI strategies cannot be overcome since the envelope features are the targets of the SE algorithms,that is,their performances are upper bounded by that of the existing standard CI strategies in quiet environments.In this thesis,a noise robust unilateral CI signal processing strategy is proposed to relieve the problems abovementioned.Neural networks(NN)are built and trained to simulate the procedure of the advanced combination encoder(ACE),a widely implemented clinical CI strategy.Not only can the proposed algorithm be compatible with the ACE processor,but also it is sophisticatedly trained with designated loss function to extract noise-robust envelopes with TFS as much as possible.The proposed algorithm is therefore called NNACE.A set of vocoder simulation experiments have been conducted to evaluate the performance of NNACE with both objective and subjective metrics.In specific,two state-of-the-art SE methods,i.e.,parametric Wiener filtering-based one and DNN-based one,are selected as two benchmarks(namely,Wiener-ACE and DNN-ACE)to compare with NNACE.The objective experimental results show that 1)NNACE significantly outperforms Wiener-ACE for the speech-shaped noise(SSN)and multi-talker babble noise(Babble)in all signal-to-noise ratio(SNR)conditions;2)NNACE performs better than DNN-ACE when the parameter sizes of both networks are comparable.Meanwhile,subjective evaluation with 10 normal-hearing participant shows that NNACE significantly outperforms ACE in terms of the speech reception thresholds(SRT),i.e.,3.91 d B lower SRT in Babble whereas 5.11 d B lower SRT in SSN.When compared with the two benchmarks,NNACE significantly outperforms Wiener-ACE while works comparably to DNN-ACE in terms of SRT improvements.In addition,the preference test tells that the speech perceptual quality of vocoded sounds from NNACE are comparable to that from ACE in quiet environments.