使用前後文相關HMM模型之國語連續語音辨認

Abstract

在本論文中，我們建立一個以隱藏式馬可夫模型為基本架構的國語連續音節辨認系統，主要考量為聲學層次上的基本音節辨認。訓練語料包含MAT2000及MAT2500，由於訓練語料十分豐富，我們考慮前後文之間的相關性，以狀態層級共享的方式用決策樹決定模型共享的依據，再以隱藏式馬可夫模型來描述語音信號。最後以計算語言學學會舉辦的語音辨認競賽所使用的測試語料來測試系統效能，所獲得的基本音節辨認率大約在70%左右。

In this thesis, the context-dependent HMM model for continuous Mandarin speech recognition system was studied. The initial-final recognition units were used in this thesis. The state splitting was used to find the context-dependent models instead of the model splitting. The state-splitting was accomplished by using acoustic decision tree method. The effectiveness of the state-splitting context-dependent HMM recognition system was confirmed by simulation on the MAT(Mandarin Across Taiwan)speech database. A HMM recognizer with 1900 state models were constructed in our study. The base-syllable recognition rate of the proposed system was 70.1% comparing with HMM using 100 final-dependent initials and 39 CI finals, 5.4% improvement was achieved.