使用右文相關聲韻母模式之國語關鍵詞辨認

Abstract

本論文首先介紹了決策樹的原理與方法，並依此方法針對國語聲母韻母的特性，產生右文相關之韻母模型作為關鍵詞辨認系統中的關鍵詞及填充模型，在辨認關鍵詞的模型相同的情況下，比較不同粗細的填充模型對關鍵詞辨認的影響，而為了減少關鍵詞中的假警報，也嘗試分析真正關鍵詞與假警報特性的不同，並在減少假警報後以關鍵詞的辨認分數挑選最有可能包含關鍵詞的路徑。最後在考慮每個音節原有特性的差異後，對其辨認分數及音框長度做正規化處理，並加入短詞效應的考量後重新挑選最佳路徑。得到最後的辨認率為86.43%，而且假警報大幅減少。

In this thesis, the context-dependent initial-final models are used to construct both the keywords and filler models in order to improving the performance of Mandarin keyword spotting system. The decision tree clustering method is used to find the right-context-dependent final models used in the system. 100 final-dependent initial models and 290 right-context-dependent final models are first constructed for the keywords. The system performance is examined when different numbers of filler models which can be found from the clustering procedure, are used in the system. In order to increasing the keyword recognition rate, the Top-N recognition result of the input sentential utterance is first found, and a verification procedure is used to filter out the candidates with lower keyword-filler likelihood ratio and shorter syllable duration. Finally, the normalized keyword recognition scores and lengthening factor are combined to find the best sentential candidate. The performance of above methods is examined in the telephone number inquiry system with 1013 keywords, the 86.4% keyword recognition rate is achieved with 0.37 FA/KW/HR false alarm rate.