不特定語者國語連續音節辨認

Abstract

論文中探討不特定語者國語連續音節之辨認，所建立的 1300 音節 辨認器，係由一個 以聲韻母模型為基礎的 411 音節隱藏式馬可夫模型 (Hidden Markov Model, HMM)辨認 器以及一個左相關的 HMM 聲調辨認 器結合而成。 為解決基音節與聲調結合時的同步問題 ，我們將每一個 音節視為由 100 類右相關聲母模型、39 類前後文不相關韻母模型以及 2 9 類右相關聲調模型所組合而成，在辨認時，先將 100 類聲母模型、39 類韻母模型以及 29 類聲調模型分別計算狀態觀測機率後，再經由查表 重組合成 1300 音節的模型，用修 正的一階動態規劃演算法進行辨認。 經電話語音測試，使用 159 人的語料訓練及 36 人 的語料測試，獲得 1300 音節的辨認率為 36.40 %， 而 411 音節的辨認率則 47.14 %。 In this thesis, a new method of continuous Mandarin syllable recognition is studied. The 1431-syllable recognizer is composed of an initial-final based 411-base- syllable recognizer and a tone recognizer. It uses 100 right-final-dependent initial HMM models, 39 context- independent final HMM models and 29 right-tone-dependent tone HMM models. To cope with the synchronization problem of combining base-syllable and tone recognizers, in the recognition process, the observation probabilities of 100-initia l models, 39-final models and 29-tone models are separately calculated first. The scores for 1431 syllables are then calculated by a table lookup process. Last, the recognition process is finished by performing a modified one- stage dynamic programming. Performance of the method was examined by simulation using a telephone speech database of 159 training speakers and 36 testing speakers. Accuracy rates of 36.40 % and 47.14 % were obtained for recognition of 1431 and 411 syllables, respectively.