以音韻屬性偵測擷取對話語音關鍵詞之研究

Study on Keyword Spotting using Prosodic Attribute Detection for Conversational Speech

在口語對話上，為了有效地理解使用者所要表達的資訊意義，擷取關鍵詞語是很重要的研究議題之一。本研究針對對話語音內容的關鍵詞，提出以音韻屬性來做為擷取關鍵詞的特徵。利用預先訓練的決策樹將語者語句分段成韻律詞，進一步使用支援向量機（SVM）來偵測韻律詞是否為關鍵詞。在此利用中研院語言研究所鄭秋豫所提出階層式多短語語流韻律架構與韻律詞邊界的偵測方法，而邊界偵測為利用其韻律特性建立的決策樹來偵測。最後，以音韻屬性為特徵來做為偵測的參數，藉由SVM分析各個韻律詞之特徵值找出焦點所在語音時間區段，藉由擷取此焦點作為關鍵詞。最後部分為針對錄製的對話語料進行實驗並分析，所得到的準確度與召回率比參照發音相似度或聲學特徵還要高，證實所提方法在關鍵詞擷取是可行的。

It is one of most essential issues to extract the keywords from conversational speech for understanding the utterances from speakers. This thesis aims at keyword spotting from spontaneous speech for keyword detecting. We proposed prosodic features that are used for keyword detection. The prosody words are segmented from speaker's utterance according to the pre-training decision tree. The supported vector machine is further used as the classifier to judge the prosody word is keyword or not. The prosody word boundary segmentation algorithm based on decision tree is illustrated. Besides the data driven feature, the knowledge obtained from the corpus observation is integrated in the decision tree. Finally, the keyword in the focus part are extracted using prosody features by sported vector machine (SVM). According to the experimental results, we can find the proposed method outperform the phone verification approach especially in recall and accuracy. This shows the proposed approach is operative for keyword detecting.