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

Denoising speech signals represent a challenging task due to the increasing number of applications and technologies currently implemented in communication and portable devices. In those applications, challenging environmental conditions such as background noise, reverberation, and other sound artifacts can affect the quality of the signals. As a result, it also impacts the systems for speech recognition, speaker identification, and sound source localization, among many others. For denoising the speech signals degraded with the many kinds and possibly different levels of noise, several algorithms have been proposed during the past decades, with recent proposals based on deep learning presented as state-of-the-art, in particular those based on Long Short-Term Memory Networks (LSTM and Bidirectional-LSMT). In this work, a comparative study on different transfer learning strategies for reducing training time and increase the effectiveness of this kind of network is presented. The reduction in training time is one of the most critical challenges due to the high computational cost of training LSTM and BLSTM. Those strategies arose from the different options to initialize the networks, using clean or noisy information of several types. Results show the convenience of transferring information from a single case of denoising network to the rest, with a significant reduction in training time and denoising capabilities of the BLSTM networks.