Combining multi-section Bayesian template with level-building algorithm for robust connected Mandarin digit recognition

A robust connected Mandarin digit recognition system based on multi-section Bayesian templates and the level-building algorithm is presented. In the multi-section Bayesian template, both coarticulation effects and the characteristics of the digit itself can be properly exhibited. In particular, each section is organized into a Bayesian template which is characterized by a powerful statistical framework. In the recognition phase, several factors such as finding a more desirable lower and upper bound of input speech length, word duration checking and post-processing are carefully considered to make the proposed system more robust. The proposed system is tested using a multi-speaker database (25 males, 25 females) and a string accuracy of 95.4% is achieved.