國語廣播新聞語音辨識之研究

Abstract

廣播新聞在現代生活裡非常普遍，結合語音辨識的技術可以應用在許多領域上。在本論文中，我們專注於廣播新聞語音辨識，首先針對廣播新聞類似自發性語音的特性，說明如何建立聲學模型，並且根據內外場環境的不同各自訓練模型，可以提高辨識效能。接著還會說明語言模型的建立，如何將一個語言模型調整成適合廣播新聞，並且使用語言模型調適的技術可以使得語言模型優化，最後的音節辨識率與基本辨識系統相比，內場可提升約16%、外場可提升20%以上。而廣播新聞的資料來源是採用中研院錄製的公視新聞資料庫（PTSND），我們也會在論文中對資料庫內容做簡介說明。

Broadcast news is very general in this day, many researches were applied by combining speech recognition technique. In the thesis, we focus on the Mandarin broadcast news speech recognition, because of the spontaneous speech characteristics, several spontaneous acoustic models were constructed, and individual model training were performed under different environments. Then, the language model construction will be described carefully. Performance will be improved by using the language model adaptation. At last, the syllable recognition rate was increased about 16% for anchor. Above 20% increase was obtained for reporter and interviewee. All related experiments proceeded over the broadcast news database：PTSND. Database content will also be introduced in this paper.