Experimental study on transfer learning in denoising autoencoders for speech enhancement

Abstract

The quality of speech signals is affected by a combination of background noise, reverberation, and other distortions in real-life environments. The processing of such signals presents important challenges for tasks such as voice or speaker recognition. To enhance signals in such challenging conditions several deep learning-based methods have been proposed. Those new methods have proven to be effective, in comparison to classical algorithms based on statistical analysis and signal processing. In particular, recurrent neural networks, especially those with long short-term memory (LSTM and BLSTM), have presented surprising results in tasks related to enhancing speech. One of the most challenging aspects of artificial neural networks is to reduce the high computational cost of the training procedure. In this work, we present a comparative study on transfer learning to accelerate and improve traditional training based on random initialization of the internal weights of the networks. The results show the advantage of the proposal in terms of less training time and better results for the task of denoising speech signals at several signal-to-noise ratio levels of white noise.