中文語音韻律架構之建立及其在語音辨認之應用

Abstract

本計畫擬探討中文韻律結構及其在中文語音辨認的應用，將以現有的統計式韻律模 式為基礎，來建構可以描述中文韻律結構之韻律片語多層架構，並討論相關議題，包括 韻律片語邊界之標示及預測、建立韻律結構和語法結構之對應模式、討論說話速度對韻 律結構的影響等；在應用韻律模式於中文語音辨認的研究中，我們擬嘗試建構包含聲學 模式、韻律模式及語言模式之語音辨認架構，將統計式基頻軌跡、能量、音長模式加入 辨認系統中，主要探討議題為韻律信息對聲學模式及語言模式之影響、如何降低系統的 複雜度等。

We shall exploit two topics in this project. One is the construction of a computational prosody structure to describe the prosody hierarchy of Mandarin speech. Issues to discuss include automatic prosodic labeling, prediction of breaks from text, the relation between prosody structure and syntactic structure, the affection of speaking rate on prosody structure, etc. Another topic is the incorporation of prosody model into automatic speech recognition. The approach of using the existing statistics-based syllable pitch contour, energy contour, and duration models to assist in Mandarin speech recognition will be explored intensively. Issues to discuss include the affections of prosodic cues on acoustic modeling and language modeling, the tradeoff between system complexity and recognition performance, etc.