一般性的關鍵詞辨識及語句驗證系統

Abstract

本論文主要在討論一般性的關鍵詞辨識及語句驗證系統。所謂的一般性可 以從兩個方面來定義：第一，我們考慮不同大小的詞彙。第二，所有的中 文內容都可以識為非關鍵詞的語句。根據前人所提出的架構，我們建立了 一個基礎系統。這個系統架構可以涵蓋所有的中文語句，而且A*演算法亦 可與此架構整合，在大字庫下達到有效率地搜尋。本論文主要討論三種非 關鍵詞模型的建構方法。以這三種非關鍵詞模型結構為基礎，我們對所提 出的非關鍵詞架構，提出了三個問題。隨後，本論文會依據這三個問題， 提出改進的方法。在使用500、5000、以及25000字的關鍵詞時，我們的基 礎系統的Top1辨識率分別為86.8%、68.6%、以及52.8%。加上我們所提出 的兩種改進方法之後，辨識率可以分別提升7%、12%以及13%。如果以每一 種非關鍵詞模型結構在500字下的最好辨識率為基礎，再加上語句驗證的 話，辨識率亦可再進一步地提昇。 This paper considers the design of a general keyword spotter for Mandarin speech and utterance verification. The design of keywords spotting is general from two aspects: First, we have considered various vocabulary size. Second all contents of Mandarin speech are assumed to be extraneous speech. We establish the baseline keyword spotting systems according to the framework of Huang et. al. All extraneous speech can be modeled under this framework. Also, the frame work can be well integrated with the tree-trellis search algorithm to achieve efficient search in large vocabulary tasks. This paper considers three varieties of filler model structures for the framework based on subsyllabic grammar of Mandarin speech. On the basis of the three structures, we infer the problem s of this framework through three arguments. This paper then presents two methods to modify the spotting mechanism according to these arguments. Thebest top 1 inclusion rates of the baseline system are 86.8%, 68.6%, and 52.8% for 500-, 5000-, and 25000-word systems. After our proposed two methods are applied, the top 1 inclusion rates can be significantly enhanced individually by about 7%, 12% and 13% for 500-, 5000-, and 25000-word systems. If taking the best results of each filler structure in 500-word vocabulary as keyword spotting system and apply utterance verification, the recognition rates can be farther improved with reasonable false rejection rate.