標題: 以修正型隱藏式半馬可夫模型做國語連續音辨認
A Modified Hidden Semi-Markov Model for continuous Mandarin ion
作者: 陳宗義
Joney Chen
陳信宏
Sin-Horng Chen
電信工程研究所
關鍵字: 修正型隱藏式半馬可夫模型;參考段;殘餘信號;modified hidden semi-Markov model;reference segment;residual
公開日期: 1994
摘要: 修正型隱藏式半馬可夫模型應用在中文獨立音節辨認上,已被證明優於傳 統的隱藏式馬可夫模型,在本論文中我們將把這種模型應用在語者相關的 中文連續音辨認上,以探測其性能。修正型隱藏式半馬可夫模型是一種音 段模型,它將每一個聲音事件,視為是由具有時變性的參考段和高斯白色 雜音信號的加成,因此修正型隱藏式半馬可夫模型比隱藏式馬可夫模型更 具有模擬非平穩性語音信號的的能力。基於修正型隱藏式半馬可夫模型之 研究,本論文內容主要可分為三部份,第一部份,以隱藏式馬可夫模型為 辨認模型,對語音資料庫做測試,其結果當作驗證修正型隱藏式半馬可夫 模型比較的依據,第二部份,建立一套適合修正型隱藏式半馬可夫模型的 自動辨認與訓練演算法,第三部份,加入語言學知識來訓練修正型隱藏式 半馬可夫模型,使能達到有限資料庫下,訓練語料共用的目的。在完成上 述研究後,我們所得到的辨認正確率最高為 80.26%。 A modified hidden semi-Markov model (MHSMM) for continuous Mandarin syllable recognition is proposed in this thesis.a conventional HMM model, it treats an acoustic segment as an event by a time-varying template with the residual being modelled by a Gaussian noise. A syllable model is then formed by concatenating acoustic segments. So, a MHSMM is a modification of the HSMM to the identical and independent observation distributions of with frame-correlatedly trend distributions. By this both the segmental k-means training algorithm and the one-stage recognition algorithm of the HMM method can be slightly modified method. Two recognition schemes are studied in this work. syllable MHSMM models while the other uses initial-final models. results showed that the latter performs slightly better. Ae of 80.26% was achieved. The performance is comparable to a initial-final based HMM method.
URI: http://140.113.39.130/cdrfb3/record/nctu/#NT830436042
http://hdl.handle.net/11536/59398
顯示於類別:畢業論文