Malay continuous speech recognition using continuous density hidden Markov model

• 主要作者: Ting, Chee Ming
• 格式: Thesis
• 語言: English
• 出版: 2007
• 主題: TK Electrical engineering. Electronics Nuclear engineering
• 在線閱讀: http://eprints.utm.my/id/eprint/6426/4/TingCheeMingMFKE2007.pdf; http://eprints.utm.my/id/eprint/6426/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:62313
• 機構: Universiti Teknologi Malaysia

This thesis describes the investigation of the use of Continuous Density Hidden Markov Model (CDHMM) for Malay Automatic Speech Recognition (ASR). The goal of this thesis is to solve the constraints of current Malay ASR that are: speaker-dependent, small vocabulary and isolated words, and provides a basis in developing speaker-independent (SI) Malay large vocabulary continuous speech recognition (LVCSR). Hidden Markov Model (HMM) based statistical modeling is used in Malay speech recognition. HMM is a robust and powerful technique capable of modeling of speech signals. With their efficient training algorithm (Baum-Welch and Viterbi/Segmental K-mean) and recognition algorithm (Viterbi), as well as it’s modeling flexibility in model topology, observation probability distribution, representation of speech unit and other knowledge sources, HMM has been successfully applied in solving various tasks in this thesis. CDHMM which model the continuous acoustic space eliminates quantization error imposed by discrete HMM. CDHMM performs better than discrete HMM in Malay speech recognition. CDHMM with mixture densities which is capable to model inter-speaker variability performs well in multi speaker task (99% in isolated words task). The result expects its well performance in SI task in the future. A connected words ASR is developed and evaluated on Malay connected digit task and has achieved reasonably good accuracy with limited training data. Segmental K-mean training procedure is proven to perform better than the manual segmentation. The sub-word unit modeling is attempted in Malay phonetic classification and segmentation on medium vocabulary Malay continuous speech database. Experiments are conducted to investigate different feature set and mixture components. The knowledge of continuous ASR architecture and sub-word unit modeling gained in this work has provided basis for Malay LVCSR. For conclusion, the basic idea of HMM implemented in other language domain can be successfully applied in the Malay language domain as well.