Throat Microphone Speech Enhancement via Progressive Learning of Spectral Mapping Based on LSTM-RNN

In this paper, we propose a progressive spectral mapping learning algorithm for throat microphone (TM) speech enhancement. Unlike previous full-band spectra mapping algorithms, this algorithm divides the spectra mapping from TM speech to Air-conducted (AC) speech into two tasks, one is the voice conversion task, and the other is the artificial bandwidth extension task. Long short-term memory recurrent neural network (LSTM-RNN) is further deployed as the mapping model. Objective evaluation results show that the TM speech quality is improved when compared with conventional full-band spectra mapping framework and DNN-based mapping model.