Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling
Gadallah, M. (1999). Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling. EKB Journal Management System, 2(2nd International Conference on Electrical Engineering ICEENG 1999), 253-263. doi: 10.21608/iceeng.1999.62508
Mahmoud E. Gadallah. "Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling". EKB Journal Management System, 2, 2nd International Conference on Electrical Engineering ICEENG 1999, 1999, 253-263. doi: 10.21608/iceeng.1999.62508
Gadallah, M. (1999). 'Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling', EKB Journal Management System, 2(2nd International Conference on Electrical Engineering ICEENG 1999), pp. 253-263. doi: 10.21608/iceeng.1999.62508
Gadallah, M. Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling. EKB Journal Management System, 1999; 2(2nd International Conference on Electrical Engineering ICEENG 1999): 253-263. doi: 10.21608/iceeng.1999.62508

Performance Improvements of Isolated Word Recognizers Based on Hidden Markov Modeling

The International Conference on Electrical Engineering
Article 28, Volume 2, 2nd International Conference on Electrical Engineering ICEENG 1999, November 1999, Page 253-263  XML PDF (2.26 MB)
Document Type: Original Article
DOI: 10.21608/iceeng.1999.62508
View on SCiNiTO View on SCiNiTO
Author
Mahmoud E. Gadallah
Assoc. Prof., Military Technical College.
Abstract
This paper introduces proposed solutions for increasing the efficiency of the isolated word recognizers (IWR) whose vocabularies comprise words with considerable difference in length. Most of the languages consist of words that have appreciable difference in the numbers of phonemes. In these cases, the recognition accuracy of the IWR systems that are based on Hidden Markov Modeling (HMM) is degraded because they usually use a fixed number of states for all the vocabulary words. These proposed approaches are originally initiated to, overcome this problem for such types of vocabularies. The proposed solutions, introduced in this work, have been developed in two stages: (i) In the first stage, the HMM is allowed to construct the words models with variable number of states (VNS-HMM) according to the length of the words. The application of this idea has shown better recognition accuracy over HMM with fixed number of states (FNS-HMM). (ii) Considerable inigrovement has been achieved in the second stage in which the reference words are categorized according to their lengths and divided into subvocabularies. Due to the applications of these approaches, two other improvements have been obtained, namely, they are: (i) the ability to increase the vocabulary size; and (ii) decreasing the recognition time.
Keywords
Hidden Markov Model (HMM); Isolated Word Recognition (IWR); Number of states of the HMM; Training phase; Recognition phase; Time compression. HMM with Fixed number of states (FNS-HMM); HMM With variable number of states (VNS-HMM)
Statistics
Article View: 116
PDF Download: 204