Hybrid Speech Enhancement with Wiener filters and Deep LSTM Denoising Autoencoders

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 network-based. In this paper, we propose a hybrid approach to speech enhancement which combines two stages: In the first stage, the well-known Wiener filter performs the task of enhancing noisy speech. In the second stage, a refinement is performed using a new multi-stream approach, which involves a collection of denoising autoencoders and auto-associative memories based on Long Short-term Memory (LSTM) networks. We carry out a comparative performance analysis using two objective measures, using artificial noise added at different signal-to-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.