片桐 滋
Relaxation-based speech labeling
Abstract:It was revealed that a trained individual, i.e., a labeler, could perform accurate speech
labeling and that such accuracy was based on his/her flexible decision process using many kinds of spectrographic features. In this paper, a new relaxation-based speech labeling system which duplicates the ability of the labelers is proposed. To realize the trial-and-error process of the labelers in the system, we have adopted a blackboard model and a discrete relaxation process. The system consists of a blackboard, and three subsystems: an acoustic analyzer, a verifier, and a supervisor. The blackboard is a working memory through which the three subsystems communicate with each other, and allows the system to realize the complicated behavior of the trial-and-error process. The acoustic analyzer computes many kinds of acoustic parameters, e.g., formant and pitch frequencies, corresponding to the spectrographic features used by the labelers. Also, the verifier is broken down into a symbol hypothesizer, and two kinds of functions: boundary detectors, and label identifiers. The verifier, with a behavior principle based on the relaxation process, efficiently performs the hypothesis verification for many of the label candidates. The supervisor controls the whole system. Preliminary experiment results show that the performance of the system is comparable to the performance of the labelers.