D特定語者之中文多音節詞語辨識

Abstract

非特定語者之中文多音節詞語辨識 學生 : 陳柏琳 指導教授 : 劉啟民 博士 國立交通大學資訊工程系碩士班 摘要 本論文從兩個主要的觀點來考量及設計一個非特定語者的中文多音節詞語 辨識 (speaker-independent Mandarin polysyllabic word recognition) 系統：語音模型的建 立與辨識速度。 我們嘗 試增加訓練資料量並且改進語音聲韻特徵的模型，以期 達 到提升辨識率的目標。 基於中文聲韻的特性，本論文設計與實驗了三種 語音 模型：context-independent INITIAL model 、 right-context-dependent INITIAL model及right-context- dependent null-INITIAL model。 最好的語音模型在500 字詞、5000字詞、25000字詞的系統，top 1詞的辨識率平均可達到99.1% 、 93.7%、83.6%，而top 3詞則平均可達到99.8% 、98.5%、95.2%。 基於上述 的辨識結果，我們考慮 利用搜尋演算法來增進搜尋的效率。 由於tree-trellis search 有能力大量縮減搜尋時所需的運算時間並且不會降低系統的辨識 率，本論 文採用此種演算法作為研究搜尋效率的基本架構 ，並且進而討論與分析四種執行 tree-trellis search 的 技巧。 為了與tree-trellis search做比較，我們更進一步發展 一 種我們稱之為fast-match search的beam search演算法。 實驗結果顯示 tree- trellis search顯著地比其他搜尋演算提供較佳的搜 尋速度，並且可以使搜尋時間 降低成幾乎與詞庫的大小無關 。 最後，我們建立了一套及時 (real-time) 的展 示系統 於Pentium-90的PC，以作為大量測試之用。 Speaker-Independent Mandarin Polysyllabic Word Recognition Student : Berlin Chen Advisor : Dr. Chi-Min Liu Institute of Computer Science and Information Engineering National Chiao Tung University Abstract This thesis considers the design of speaker-independent Mandarin polysyllabic word recognition system from two main viewpoints: the phonetic modeling and the recognition speeds. We first establish a baseline system based on a system established last year. The baseline system improves the recognition performance by increasing the training data and adopting another feature. On the baseline system, we consider the acoustic characteristics of Mandarin speech for phonetic modeling. We design and experiment with three phonetic models: context-independent INITIAL model, right- context-dependent INITAL model, and right-context-dependent null-INITIAL model. For the most accurate model, in the 500-word, 5000-word, and 25000-word tasks, the system can provide respectively an average recognition rate 99.1%, 93.7% and 83.6% for top 1 word, and 99.8%, 98.5% and 95.2% for top 3 words. On the basis of the recognition results, we consider various search algorithms to enhance the computing speeds. Since that the tree-trellis search has the potential to greatly reduce the computation time without deteriorating the recognition rate of the baseline system, we adopt the search algorithm as the basic framework and investigate four implementation techniques. In comparison with the tree-trellis search, we further develop a kind of beam search algorithm, which we call the fast- match search. The experimental results show that the tree-trellis search provides significantly better speeds than other search algorithms, and has a search time only slightly dependent with word size. A real-time demo system has been implemented on the Pentium-90 PC for vast testing.