Hybrid Speech Enhancement with Wiener filters and Deep LSTM Denoising Autoencoders

JOHN GODDARD CLOSE; MARVIN COTO-JIMENEZ; HUGO LEONARDO RUFINER; LEANDRO EZEQUIEL DI PERSIA

San Carlos

Conferencia; 218 IEEE International Work Conference on Bioinspired Intelligence; 2018

IEEE - Universidad de Costa Rica- Técnológico de Costa Rica

Over the past several decades, numerous speech enhancement techniques have been proposed to improve the performance of modern communication devices in noisy environments. Among them, there is a large range of classical algorithms (e.g. spectral subtraction, Wiener filtering and Bayesian-based enhancement), and more recently several deep neural networkbased. In this paper, we propose a hybrid approach to speechenhancement which combines two stages: In the first stage, thewell-known Wiener filter performs the task of enhancing noisyspeech. In the second stage, a refinement is performed usinga new multi-stream approach, which involves a collection ofdenoising autoencoders and auto-associative memories based onLong Short-term Memory (LSTM) networks. We carry out a comparative performance analysis using two objective measures, using artificial noise added at different signalto-noise levels. Results show that this hybrid system improves the signal?s enhancement significantly in comparison to the Wiener filtering and the LSTM networks separately.