Book Description

Many types of acoustic-phonetic constraints can be applied in speech recognition. Shipman and Zue proposed an isolated word recognition model in which sequential constraints are applied at a broad phonetic level to hypothesize word candidates. Detailed acoustic constraints are then applied on a subsequent phone representation to determine the best word from the remaining word candidates. This thesis examines how their model can be extended to continuous speech. We used the recognition of continuously spoken digits as a case study. We first conducted a feasibility study in which words and word boundaries were hypothesized from an ideal broad phonetic representation of a digit string. We found that strong sequential constraints exist in continuous digit strings and used these results to extend the Shipman and Zue isolated word recognition model to continuous speech. The continuous speech model consists of three components: broad phonetic classifier, lexical component, and verifier. These components have been implemented for the digit vocabulary for the purpose of exploring how acoustic-phonetic constraints can be applied to natural speech. The broad phonetic classifier produces a string of broad phonetic labels from a set of parameters describing the speech signal. The lexical component uses knowledge about statistical characteristics of the output produced by the broad phonetic classifier to score each of the word hypothesis. Evaluation of this part of the system suggests that it can prune unlikely word candidates effectively. Nine acoustic features were defined to characterize phones for verifying each of the word candidates. Evaluation of the verifier on the digit vocabulary demonstrates the power of a phone-based representation and of using a few well-motivated acoustic features for describing phones in an acoustic-phonetic approach.