Please use this identifier to cite or link to this item:
標題: 隱藏式馬可夫模型於聲控選單環境控制系統之應用
The Application of Hidden Markov Model to a Menu-Driven Environmental Control System.
作者: 黃朝元
Yuan, Huang Chao
關鍵字: left-to-right;前後型
出版社: 電機工程學系
本論文研究利用以MFCC(Mel Frequency Cepstrum Computation)汲取出語音的特徵參數,以連續型隱藏式馬可夫模型(Continuous Observation Hidden Markov Model )為主的語音辨識演算法,應用連續型隱藏式馬可夫模型,分別建立各個語音的模型。在隱性馬可夫模型上,使用"前後型"的模型,並且在計算機率時引用連續性機率的函數,訓練時,k-means分割疊代程序提供良好的初始估測,使用疊代的方法估測每個數字模型,因為疊代會使新的模型更接近這個數字的最佳模型,辨識時計算所有模型機率找出最大結果即是。研究以單一發語者為實驗,實驗不同樣本數,不同混合數和不同狀態數對系統辯識影響。實驗結果顯示在使用狀態數5、混合數3和樣本數70下,可達99.3%的辨識率。

In this thesis, a menu-driven environmental control system is designed for the individuals with disabilities. According to the number of different sounds (0-9) that the user can utter, the menu-driven system is designed to facilitate the selection of the desired function provided in the environmental control system to improve the recreation ability, the functionality, and the independence of the individual with disability.
Speech feature parameter extraction by using mel frequency cepstrum computation (MFCC) technology is studied. The continuous observation hidden markov model (HMM) is used as the basis of speech recognition system. In the model topology, the left-to-right speech model and the continuous probability density function are adapted. In the training, segmental k-means segmentation with clustering provided good initial estimates, each number's mode parameter was reestimated by iteration procedure, new mode parameter must be close to optimal mode parameter. In the recognition procedure, computing all mode was to find the best result of probability in HMM. The study is based on the single speaker with different samples, states and mixtures to test the effectiveness of the recognition system. Experimental results show that the system can achieve recognition rate 99.3% using five states, three mixtures and seventy samples for each class.
Appears in Collections:電機工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.