Towards a Neural-Based Single Channel Speech Enhancement Model for Hearing-Aids

Advancements in machine learning techniques have promoted the use of deep neural networks (DNNs) forsupervised speech enhancement. However, the DNN’s benefits of non-explicit noise statistics and nonlinearmodeling capacity come at the expense of increased computational complexity for training and inference whichis an issue for real-time restricted applications, like hearing aids. Contrary to the conventional approach whichseparately models the feature extraction and temporal dependency through a sequence of convolutional layersfollowed by a fully-connected recurrent layer, this work promotes the use of convolutional recurrent network lay-ers for single-channel speech enhancement. Thereby, temporal correlations among inherently extracted spectralfeature vectors are exploited, while further reducing the parameter set to be estimated relative to the conven-tional method. The proposed method is compared to a recent low algorithmic delay architecture. The modelswere trained in a speaker independent fashion on the NSDTSEA data set composed of different environmen-tal noises. While objective speech quality and intelligibility measures of the two architectures are similar, thenumber of network parameters in the suggested enhancement method being reduced by 66%. This reduction ishighly beneficial for storage and computation constraint applications