漢語朗讀語音停頓產生規則之初步分析

Abstract

本論文針對漢語朗讀語音的停頓形成規則做初步統計分析，由於在文字轉語音的系統中沒有聲學參數的輔助，只能利用文字分析器得到的語言參數來預估停頓韻律標記，一般來說文字分析器得到的語言參數大多是詞層次以及語句層次的語言參數，但這些參數對於停頓韻律的預估仍不太足夠。要增進中文語音停頓韻律的預估效能，使合成出來的中文語音更加自然，還需要更豐富的語法結構及語意層次的語言參數。本研究以統計的方式分析特殊詞類的前後及片語內特殊位置的音節邊界的停頓分布，所探討的詞類包括介詞及動作句賓動詞，也將相鄰關係密切的兩種詞類做合併再做前後停頓分布的討論，並探討詞的語法功能與停頓的關係；另外也分析成語邊界的停頓分布。研究結果發現，相鄰詞的詞性、特殊片語的結構、和詞本身的語法功能對於特殊詞類的詞邊界停頓有一定程度的影響，片語內特殊位置的停頓也會受到片語結構的影響，這些特性可以形成有意義的規則，來增進由語言參數做停頓預估的效能。

A preliminary study on break formation rules for Mandarin read speech is reported in this thesis. In Mandarin text-to-speech, break prediction is conducted using linguistic features obtained by analyzing input text without the help of prosodic-acoustic features. Its performance is usually fair because of using only word- and sentence-level features. More sophisticated syntactic or semantic features are needed to improve its performance. This study explores the distributions of 7 break types for boundaries of some special words, word chunks, and phrases. Specifically, word chunks or phrases formed by prepositions or VEs are intensively studied. Besides, beak-type distributions for boundaries of idioms are also discussed. We find that some features are highly related to juncture’s break type, including the POS of neighboring word, phrase structure, and pragmatic function of word. Some break formation rules are hence concluded. They are expected to be useful for break prediction in the future study.