| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 94 === Basically, the state-of-the-art automatic speech recognition (ASR) systems are based on techniques of dynamic programming and hidden Markov model. There are several crucial issues happening in building desirable ASR performance. Among them, how to reliably detect change points of continuous speech in presence of high co-articulation effect and distortion environments plays a critical role. In the literature, likelihood-ratio (LR) based confidence measure was developed to improve detection performance. This likelihood ratio (LR) criterion could be used to decide the acceptance or rejection for the alignment between speech frames and acoustic models/units. However, in case of spontaneous-style speech, the probabilistic scores in some intervals turn out to be vibrating and confusing. This causes unreliable alignment during search processing for large vocabulary continuous speech recognition (LVCSR). Previously, some methods were presented to detect change points in HMM state level. But, these works should specify empirical detection threshold and were not considered as a direct solution to overcome vibration problems in boundaries of speech units.
In this thesis, we present the run test approach to test the randomness of the states of decision probabilistic scores in observation speech sequence. The non-parametric statistics is calculated and used to determine the optimal change point with the best randomness for the states before and after the change point. Through combining this principle and LR criterion, we can sequentially detect change points for building desirable LVCSR search algorithm. In the experiments, we implement and evaluate this approach using TDT2 Mandarin broadcast news corpus.
|
|---|
