Auditory filterbank denoising neural network for speech enhancement in wearable auditory device

In this study, a speech enhancing neural network (NN) is proposed, which is designed for monaural auditory devices, specifically designed for use in hearing aids. Herein, a 32-channel auditory filterbank (FB) is first implemented with an algorithm processing delay of 8 ms, which is tailored to meet the requirements of auditory devices. The proposed method primarily aims to integrate a denoising NN within the analysis phase of a uniform polyphase discrete Fourier transform (DFT) FB, aimed at enhancing speech within each band. For the denoising model, complex-valued convolutional NNs have been applied, specifically targeting the restoration of speech phase information based on the spectral components of the DFT. A multi-loss method is introduced, which is designed to further account for the loss of analysed speech signals within the split bands during the training process, leveraging the DFT FB strategy. To evaluate the efficacy of the proposed method, objective assessments of speech intelligibility and quality scores are conducted under various noise conditions. The results demonstrate that the proposed method can outperform the existing method across all types of noise. The proposed auditory filterbank denoising neural network aims at enhancing speech within each band by integrating a denoising neural network within the analysis phase of a uniform polyphase discrete Fourier transform filterbank for auditory devices such as hearing aids. All components of the proposed architecture, that is, analysis filterbank, synthesis filterbank and speech denoising model, are integrated into a single neural network architecture, and used for inference and training. image