Title :
A study of training strategy of unit HMM for Chinese speech recognition
Author :
Xiaobo, Pi ; Limin, Du ; Ziqiang, Hou
Author_Institution :
Inst. of Acoust., Acad. Sinica, Beijing, China
Abstract :
In an automatic speech recognition system based on HMM, the training of the unit HMM and the sequence of unit HMMs is not globally optimized therefore, the choice of initial parameters of the unit HMM, and the controlling for evolution of the training are crucial to the final results of the training and recognition accuracy. Rational considerations of the strategy for the training of the unit HMM and sequence of unit HMMs should be beneficial to performance improvement. We study the issue by comparing several methods of initialization of the parameters of the unit HMM, the controlling strategies of the training process and their effects on the recognition accuracy
Keywords :
hidden Markov models; natural languages; speech recognition; Chinese speech recognition; automatic speech recognition system; initial parameters; performance improvement; recognition accuracy; training process; training strategy; unit HMM; Acoustics; Automatic speech recognition; Covariance matrix; Design optimization; Gaussian processes; Hidden Markov models; Process control; Speech recognition; Testing; Training data;
Conference_Titel :
Signal Processing, 1996., 3rd International Conference on
Conference_Location :
Beijing
Print_ISBN :
0-7803-2912-0
DOI :
10.1109/ICSIGP.1996.567378
