Application of neural network speech dereverberation for cochlear implants

This study aims to evaluate different neural network models, including HELM, DDAE, IDEA, LS-LSTM, and Late Suppression Unet (LS-Unet), on normal speech and speech for cochlear implant in various environments involving reverberation and noise. The TMHINT dataset was used for experimentation, and objective evaluation methods employed the Short-Time Objective Intelligibility (STOI) and the Normalized Covariance Metric (NCM). The results from experiments on normal speech demonstrated that LS-Unet achieves the greatest improvement in Chinese speech intelligibility in noisy and reverberant environments, respectively. In experiments involving noise and reverberation, training models with noisy and reverberant speech does not improve speech intelligibility when tested on speech containing both noise and reverberation. However, the LS-Unet (Noise) model, trained on noisy speech, proved effective in enhancing the intelligibility of speech containing both noise and reverberation. When we compared LS-Unet (Noise) combined with the ElectrodeNet-CS strategy and LS-Unet (Noise) combined with the ACE strategy, we can observe that the average NCM evaluation result for the ACE strategy is slightly better than that of the ElectrodeNet-CS strategy. However, the difference is not significant, which suggests that the deep learning-based ElectrodeNet-CS strategy has the potential to surpass the traditional ACE strategy.