國語廣播新聞語音基本辨認系統之建立

Abstract

廣播新聞節目在生活中相當普遍，而語音辨識技術在這方面有許多相關應用值得發展。在本論文中，我們首先利用廣播新聞語料庫（MATBN）依據語者環境個別建立聲學模型，並利用其中自發性語音現象，訓練出相對應的聲學模型，作為基本辨識系統。接下來，建立出語言模型並經過調適，使其特性更符合廣播新聞環境中使用，另外，為了將標點符號的訊息也加入辨識系統中，我們在建立語言模型的過程中額外保留了標點符號之相關轉移機率，並且建立出音節間靜音長度模型配合使用提升辨識率。最後，還將更上層的詞類資訊也加入使用，以求辨識器效能的進一步提升。將最終的音節辨識率與基本辨識系統相比，內場主播、外場記者與外場受訪者的辨識率各約提昇16.04%、20.52%與22.79%，此外，我們還將觀察標點符號自動標識所得到的結果。

Broadcast news (BN) is very general in our daily life, many researches were investigated into combining speech recognition technique. In the thesis, the Mandarin broadcast news corpus, MATBN, was adopted to train the acoustic model at first. And the models for spontaneous phenomena were built by making use of this corpus, too. Then, the language model (LM) for BN recognition system was built and adapted. Besides, in order to exploit the information of punctuation mark, the related transition probabilities were retained while building LM, and the pause duration models were trained to improve the performance. At last, we tried to use the information on part of speech at the same time. Finally, the syllable recognition rate was increased about 16.04% for anchor, 20.52% for reporter and 22.79% for interviewee. In addition to the recognition rate, the performance of automatic punctuation will be observed and analyzed.