A New Framework for Mandarin Lvcsr Based on One-pass Decoder

This paper describes a new framework based on one-pass and decision tree based class-triphone acoustic modeling for Mandarin LVCSR. Compared with the multi-pass decoder, it should be more knowledgeable and efficient as all sources are used at the same time when the decoder could be well organized and optimized. We give a detail about the organization of our one-pass decoder and how to handle the search space explosion by giant number of triphone and cross-word extension dealing with unknown right context including the tone context. The experimental results show that the character error rate (CER) was reduced to 13.04% for open LM and 2.8% for close LM with non-tonal class-triphone model based on the male test database from China National Hi-Tech Project 863. And with tonal class-triphone model, CER reaches 10.31% and has a 21% relative character error reduction compared with non-tonal class-triphone model.