Rule-based triphone mapping for acoustic modeling in automatic speech recognition

Abstract

This paper presents rule-based triphone mapping for acoustic models training in automatic speech recognition. We test if the incorporation of expanded knowledge at the level of parameter tying in acoustic modeling improves the performance of automatic speech recognition in Slovak. We propose a novel technique of knowledge-based triphone tying, which allows the synthesis of unseen triphones. The proposed technique is compared with decision tree-based state tying, and it is shown that for bigger acoustic models, at a size of 3000 states and more, a triphone mapped HMM system achieves better performance than a tree-based state tying system on a large vocabulary continuous speech transription task. Experiments, performed using 350 hours of a Slovak audio database of mixed read and spontaneous speech, are presented. Relative decrease of word error rate was 4.23% for models with 7500 states, and 4.13% at 11500 states. © 2011 Springer-Verlag.