Speaker-independent Mandarin polysyllabic word recognition

In this paper, we consider the design of speaker-independent Mandarin polysyllabic word recognition system from two viewpoints: the phonetical modeling and the recognition speeds. For phonetical modeling, we consider the accurate acoustic models that can increase the recognition rate. This paper experiments three phonetical models: context-independent INITIALs, right-context-dependent null-INITIALs models, and right-context-dependent INITIALs and null-INITIALs models. The recognition results show respectively an average recognition rate 99.1%, 93.7% and 83.6% for 500-words, 5000-words, and 25000-words tasks. While for top 3 words, the average rates 99.8%, 98.5% and 95.2% are achieved. On the basis of the recognition results, we consider the fast computing algorithms to increase computing speeds. Since that the tree-trellis search algorithm can retain the recognition rate of system and has the potential to greatly reduce the search time, this paper adopts the search algorithm as the basic framework and investigates some implementation techniques. The results show that the tree-trellis algorithm can provide a search time slightly dependent with word size.