Speech Enhancement Using Hybrid Model with Cochleagram Speech Feature

Abstract

The present work proposes a robust Cochleagram based speech enhancement method using a hybrid model, which is a combination of Denoising Autoencoder (DAE) and Long Short Term Memory (LSTM). The denoising autoencoder model is used for noise removal, but it lacks temporal dependency while the LSTM model shows good temporal connectivity, hence in this work, a hybrid model is proposed which helps to improve the sequential dependence as well as denoising the noisy speech signal. The performance of the proposed Hybrid model is evaluated with unknown noises, and the results of the proposed model are compared with the existing basic Denoising autoencoder and LSTM based speech enhancement algorithms. The proposed model shows improved intelligibility in both low and high signal-to-noise ratio (SNR) environments; higher intelligibility is observed in the negative SNR conditions, whereas the quality results show a slight improvement at all SNR values.