Further results on computer recognition on nonverbal speech using enhanced hidden Markov modeling

摘要

To improve the performance of computer recognition of nonverbal speech, an enhanced procedure is developed. A case study is reported to test performance. Experimental data indicate that the performance of recognition with the clipping approach is improved compared with the normal HMM approach, apparently due to the fact that the clipping procedure eliminates some irregularity in CP speech by assigning a unique VQ symbol in transition regions. This process relatively increases statistical information about 'uncertainty' in speech and reduces variability of speech. However, the obvious drawback of this approach is that for large vocabularies, ambiguity is likely to result among phonetically similar words.