A Novel Unified Framework for Speech Enhancement and Bandwidth Extension Based on Jointly Trained Neural Networks
Bin Liu, Jianhua Tao, Yibin Zheng · 2018
In this paper, we propose a unified framework for speech enhancement and bandwidth extension. The speech bandwidth extension (BWE) is investigated in noisy environment. Firstly, a Bidirectional Long Short-Term Memory Recurrent Neural Networks (BLSTM-RNN) is trained to map the noisy to clean speech features. Secondly, the BWE is also a BLSTM-RNN model. The feature enhancement neural network serves as a noise normalization module which aimed at explicitly generating the clean features which are easier to BWE by the following neural network. We combined Griffin-Lim algorithm with proposed jointly model to reconstruct wideband speech. To reduce the size of model while maintaining a similar performance, multi-task transfer learning solution is proposed. Experimental results demonstrate that the proposed framework can achieve significant improvements in both objective and subjective measures over the different baseline methods.