Evaluation of Mixed Deep Neural Networks for Reverberant Speech Enhancement

Abstract

Speech signals are degraded in real-life environments, product of background noise or other factors. The processing of such signals for voice recognition and voice analysis systems presents important challenges. One of the conditions that make adverse quality difficult to handle in those systems is reverberation, produced by sound wave reflections that travel from the source to the microphone in multiple directions.To enhance signals in such adverse conditions, several deep learning-based methods have been proposed and proven to be effective. Recently, recurrent neural networks, especially those with long and short-term memory (LSTM), have presented surprising results in tasks related to time-dependent processing of signals, such as speech. One of the most challenging aspects of LSTM networks is the high computational cost of the training procedure, which has limited extended experimentation in several cases. In this work, we present a proposal to evaluate the hybrid models of neural networks to learn different reverberation conditions without any previous information. The results show that some combination of LSTM and perceptron layers produce good results in comparison to those from pure LSTM networks, given a fixed number of layers. The evaluation has been made based on quality measurements of the signal's spectrum, training time of the networks and statistical validation of results. Results help to affirm the fact that hybrid networks represent an important solution for speech signal enhancement, with advantages in efficiency, but without a significan drop in quality.

Assessing the effectiveness of transfer learning strategies in BLSTM networks for speech fenoising

“…One example of this successful method of noise reduction is the Long Short-term Memory (LSTM) neural networks and its bidirectional extension (BLSTM), which are models of recurrent neural networks (RNNs) [8]. Particularly in speech recognition, LSTM has shown better results than DNN or convolutional networks [9,10].…”

“…On the order hand, their training procedures represent a high computational cost. For this reason, a study presented in [8] explained the advantages of using mixed neural networks for reducing computational cost in the task of reverberant speech enhancement. In this work, a comparative study is presented on different transfer learning strategies to improve the capacity of BLSTM neural networks for noise reduction and reducing training time in a set of different noise types and levels.…”