Coarse classification using a hierarchical decision tree and top down parsing

摘要: In this paper, we describe a robust technique for segmenting an utterance into sequence of coarse phonetic classes. The resulting class string is used to provide contextual information further analysis, and in lexical access limit the number word candidates. Each 10 ms interval first given probability belonging each five classes: silence, vowel, nasal-like, strong fricative weak fricative. probabilities are assigned using hierarchical classification scheme with Gaussian classifiers at node. A fuzzy C-means clustering procedure learn means variances from unlabeled data. Dynamic programming align all possible strings lexicon. performance classifier has been evaluated on TI speaker independent isolated digits correct hypothesized by more than 99 percent time.