中文自發性語音辨認系統

Abstract

近年來朗讀式語音辨認技術已經相當成熟，下一階段的目標將轉往自發性語音辨認，因此本論文將建立一套中文自發性語音辨認系統。在本論文中首先將說明如何建立自發性語音聲學模型；在語言模型方面，自發性語音中有許多常用的口語詞、感嘆詞或語助詞，這些都與傳統語言模型差異甚大，因此本論文以傳統語言模型為基礎，建立一套自發性語音語言模型。 在自發性語音韻律模型的改進方面，本研究將重新定義自發性語音中的特殊韻律現象。除此之外，傳統語音辨認通常只使用聲學模型及語言模型，但是自發性語音中有許多的特殊現象是朗讀式語音中沒有的，因此本論文試著將韻律模型也運用進來，希望能利用韻律資訊來解決這些狀況。

In recent years, the read-speech recognition technology has been quite mature, and the goal for the next phase will be transferred to spontaneous speech recognition; therefore, this paper will establish a Chinese spontaneous speech recognition system. In this paper, we will describe how to build the spontaneous acoustic model first. In the language model, there are many popular spoken word, particle and expletive in the spontaneous speech, which has great difference with the traditional language model. So, we will adapt a spontaneous speech language model based on the traditional language model in this paper. In the improvement of the spontaneous speech prosody model, this paper will re-define the special prosodic phenomena of the spontaneous speech. In addition, people usually use the acoustic model and the language model in the traditional speech recognition. However, there are many differences between read-speech and spontaneous speech. So, this paper will try to use some prosodic information to help the speech recognition and wish to resolve these situations.