Towards Robustness to Speech Rate in Mandarin All-Syllable Recognition

In mandarin all-syllable recognition, many insert errors occur due to theinfluence of non-consonant syllables. Introducing the duration modelinto the recognition process is a direct way to lessen these errors.But that usually could not work well as expected, for the durationis sensitive to speech rate. Hence, aiming at this problem, a novelcontext dependent duration distribution normalized by speech rate isproposed in this paper and applied to a speech recognition system basedon the frame of improved Hidden Markov Model (HMM). To realize thisalgorithm, the authors employ a new method to estimate the speech rate of asentence; then compute the duration probability combined with speechrate; and finally implement this duration information in thepost-processing stage. With little change in the recognition processand resource demand, the duration model is adopted efficiently in thesystem. The experimental results indicate that the syllable error ratesdecrease significantly in two different speech corpora. Especially forthe insertions, the error rates reduce about sixty to eighty percent.