SPEAKER INDEPENDENT ACOUSTIC MODELING FOR LARGE VOCABULARY BI-LINGUAL TAIWANESE/MANDARIN CONTINUOUS SPEECH RECOGNITION
Dau-Cheng Lyu, Bo-Hou Yang, Min-Siong Liang, Ren-Yuan Lyu, Chun‐Nan Hsu · 2002
In this paper, we describe the acoustic modelling technique for a bi-lingual Taiwanese /Mandarin speech recognition system, which deals with speaker independent continuous speech based on HMMs clustered by an acoustic phonetic decision tree. A bi-lingual recogniser with a bi- lingual database of 120 people was built. The vocabulary size of this system is up to 40 thousands. Unigram, bi-gram, and tree lexicon language models have been used. In order to share the common part of the two languages, a decision-tree based clustering is adopted in inter-syllable triphone units. A 89.8% word accuracy is achieved by searching on a tree lexicon net under a context free grammar.