A large vocabulary speech recognition system for Turkish

This thesis presents a large vocabulary isolated word speech recognition system for Turkish. The triphones modeled by three-state Hidden Markov Models (HMM) are used as the smallest unit for the recognition. The HMM model of a word is constructed by using the HMM models of the triphones which make up the word. In the training stage, the word model is trained as a whole and then each HMM model of the triphones is extracted from the word model and it is stored individually. In the recognition stage, HMM models of triphones are used to construct the HMM models of the words in the dictionary. In this way, the words that are not trained can be recognized in the recognition stage. A new dictionary model based on trie structure is introduced for Turkish with a new search strategy for a given word. This search strategy performs breadth-first traversal on the trie and uses the appropriate region of the speech signal at each level of the trie. Moreover, it is integrated with a pruning strategy to improve both the system response time and recognition rate.