Mandarin Digits Recognition based upon the Segmental Probability Models with Automatic Boundary Labeling

M.-T. Lin and H.-P. Chen (Taiwan)


Segmental Probability Model, Mandarin Speech recognition, Speech/nonspeech Classification, Boundary Detection, and Digit Speech Recognition.


In this paper, we successfully developed a fast scheme based on segmental probability models for Mandarin digit speech recognition with a new system for the automatic segmentation and labeling of Mandarin digit speech. The system is capable of labeling speech generated without requiring extensive linguistic knowledge or large training databases. In this work, a new approach using likelihood ratio test was used to design a speech/nonspeech classification. Further discrimination enables the identification of all Mandarin digit speech employing continuous density hidden Markov model (CHMM) and segmental probability model (SPM) alternatively. Experimental results showed that this approach could provide correct classification and significant improvements in the recognition perform The complete system achieved accuracy percentage of 93.8% no matter the SPM or the CHMM was used in the second stage, but the SPM had the advantage of requiring only about 1/8 processing time.

Important Links:

Go Back