Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

Research on Speech Units Modeling in Continuous Speech Recognition

Xiaodong He, Jian Liu, Jianlai Zhou, Tiecheng Yu

Speech Processing Laboratory, Institute of Acoustics, Chinese Academy of Sciences, Beijing, China

It is often expedient to consider using more than one single HMM to characterize a speech unit. In this paper, we suggest a new speech units modeling method based on analysis of parameters of HMMs obtained by preliminary training. By analyzing the emission probability function of a state of a HMM obtained by segmental k-means training, we can obtain the distribution of the source data and determine the splitting of that model. The experimental results, based on totally 264,500 phoneme occurring in the 9180 sentences from 60 speakers, showed that approximate 10% improvement of the recognition rate of the basic phoneme was achieved.

Full Paper (PDF)   Gnu-Zipped Postscript

Bibliographic reference.  He, Xiaodong / Liu, Jian / Zhou, Jianlai / Yu, Tiecheng (1999): "Research on speech units modeling in continuous speech recognition", In EUROSPEECH'99, 1319-1322.