In this paper, prosodic information, a very special and important feature of Mandarin speech, is used for Mandarin word verification. A two-stage strategy, with recognition followed by verification, is adopted in word recognition of telephone speech. For word recognition, 94 right context-dependent INITIALs, 37 context-independent FINALs in Mandarin speech and one silence model are used as the basic recognition units. For word verification, 15 anti-subsyllable Hidden Markov Models (HMMs), 175 context-dependent prosodic HMM's, and five anti-prosodic HMM's, are constructed. A word verification function combining phonemic-phase and prosodic-phase verification is investigated. Using a test set of 2200 word utterances from 22 speakers (14 males and 8 females), at 1.9% false rejection, the proposed verification method obtained a 9.0% false alarm rate. Comparison with a baseline system without prosodic-phase verification shows that, using prosodic information, the proposed system yields a false alarm rate reduction of 22.5% and a false rejection rate reduction of 20.8%, respectively.
|Number of pages||7|
|Journal||Proceedings of the National Science Council, Republic of China, Part A: Physical Science and Engineering|
|Publication status||Published - 1999 Jul 1|
All Science Journal Classification (ASJC) codes