HMM/ANN System for Vietnamese Continuous Digit Recognition

The study of a system for Vietnamese continuous digit recognition is described. The CSLU Toolkit was used to develop and implement hybrid HMM/ANN recognition systems. Experiments were done with a corpus of 442 sentences with 2340 words, which were extracted from two telephone-speech corpora: “22 Language v1.2” and “Multi-Language Telephone Speech v1.2”. In our experiments, a context-dependent phoneme recognizer has achieved better recognition performance than a context-dependent demi-syllable recognizer and a context-independent phoneme recognizer. Among feature sets applied to the context-dependent phoneme recognizer, the set of 12 PLP features with CMS, energy and corresponding delta values has achieved the best recognition result (96.83% word accuracy and 87.67% sentence correct).