國語韻律訊息之偵測及應用

Abstract

In this thesis, a method to detect the prosodic states of speech signals is proposed. It first employs an RNN to discriminate each input frame of an input utterance among three broad classes of syllable initial, syllable final, and silence. Outputs of the RNN are then used to drive an FSM to segment the input utterance into segments of four states. They include three stable states of I (initial), F (final), and S (silence), and a transient state of T (transition). Several acoustic cues are then extracted from the vicinities of final segments, and used to model the prosodic states of inter-final- segment periods. Two prosodic-state modeling schemes are studied. One uses VQ to directly classify the acoustic cues of two contiguous final segments into 8 or 16 prosodic states. The other uses an RNN with some linguistic features as target outputs. Prosodic states are obtained by vector- quantizing the outputs of the hidden layer of the RNN. Linguistically meaningful interpretations of these prosodic states can be observed. Finally, two outputs of the RNN , which provide word-boundary cues, are integrated into an MRNN-based continuous Mandarin word recognizer. Experimental results showed that it is helpful in improving the word recognition performance.