An RNN-based channel classification for Mandarin speech recognition over GSM/PSTN transmission environments

This paper is concerned with adopting an RNN (Recurrent Neural Network)-based channel classification technique for improving the robustness of speech recognition over GSM (Global System for Mobile Communication) and PSTN (Public Switched Telephone Network) transmission channels. We apply the RNN-based channel classification to select a most likely HMM from pre-trained HMMs that are trained for each specific channel environment. A broad-class discrimination is incorporated into the RNN-based channel classification by rejecting the disturbed frames of testing speech for improving the performance. By applying the proposed technique we obtained a drop on the average word error rate by about 24% for the recognition of the abbreviated Taiwan stock names over the conventional HMM-based scheme. Experimental results show it is an efficient framework to enhance the robustness across different channel environments.