國客雙語語音辨認

Abstract

本論文進行客語與國客雙語的語音辨認研究，重點在於如何在極有限的客語文字資料限制下，訓練一個較可靠的語言模型。在客語語音辨認上，我們首先使用客語文字資料直接訓練出一個簡單的語言模型，接著使用詞類資訊(part of speech, POS)及國客語之間的詞條對譯資訊來協助改善客語語言模型。在雙語的語音辨認上，我們嘗詴兩種方法來產生雙語聲學模型，一種是直接將國語及客語的聲學模型合併，另一種是使用相似度量測來定義音素間的距離，用以合併國客語音素成一個共用的音素集，再訓練出一個混合的雙語聲學模型。實驗結果顯示我們所提出的聲學模型與語言模型對於客語及國客雙語語音辨認效能皆有所改進。

A first study on Hakka and mixed Hakka-Mandarin speech recognition (SR) is reported in this thesis. The main focus of the study is on solving the problem of the lack of a large text corpus for training a reliable language model. In the Hakka SR, several methods to use the information of part of speech and Hakka-Chinese word translation to assist in language modeling are proposed. For mixed language SR, a method to train a mixed Hakka-Mandarin acoustic model is suggested. Experimental results show that the proposed language and acoustic modeling approaches are promising for Hakka and mixed Hakka-Mandarin SR.