本研究使用有限狀態機技術，將語音辨識所需要的聲學模型、字典和語言模型等三種知識來源看成不同階層的對應關係。透過結合，可以將分散的知識來源整合在單一個搜尋空間裡，同時保有各個階層完整的資訊。這個概念是讓搜尋空間涵蓋所有資訊，並且單純扮演引導辨識器的角色，字詞的關係與對應的發音都在離線階段處理完畢，辨識器只要負責計算即可。然而將字詞與發音的對應關係全部展開，雖然辨識過程比較簡單，但是卻存在許多相同的路徑，因此我們進一步提出部分展開的概念：根據硬體要求，調整搜尋空間部分路徑的收合，實驗結果證明省下的記憶體可達到40%。這個彈性的作法，可因應各式行動裝置或資源極有限的系統，在時間和空間上可以取得最佳的平衡，並且不受限於特殊平台或硬體。

This study uses Finite State Machine technology in speech recognition. It considers three knowledge sources in typical speech recognition: acoustic models, dictionary and language model as graphs. Each carries the mapping of different level's information. With composition, the information can be combined into a single search space, providing the way for decoder to traverse. However, full-expansion search space occupies more memory since some of the paths are the same. A partial-expansion method is thus proposed to deal with this issue, where the expansion ratio is adjusted in accordance with device's resource. Experimental results demonstrate that this method significantly reduces memory usage up to 40%, retaining the same recognition rate without consuming more computational power.