A comparison of word and syllable-based speech recognition systems

In this thesis, word and syllable-based Turkish speech recognition systems developed and compared. The developed systems are discrete, middle-sized and user-dependent. In these systems, Dynamic Time Warping (DTW), MultiLayer Perceptron (MLP), Support Vector Machine (SVM) and Hidden Markov Model (HMM) methods are used in training and testing operations. Using HMM, SVM and MLP methods, word and syllable models are generated for every word and syllable. The recognition operation is applied with this models. The developed systems consist of preprocessing, feature extraction, word and syllable training, recognition and postprocessing operations. Postprocessing has been implemented for syllable-based systems because these approach can be applied for subwords as syllables or letters. In preprocessing, speech signals are flattened, and windowing has been made before the boundaries of the word and syllables are detected. In feature extraction phase, the vectors of Mel Frequency Cepstral Coefficient (MFCC) features which represent the utterances of word and syllables are constructed for each word and syllable. After these feature vectors are trained by HMM, MLP and SVM, the recognition operation of each word in test set is made to measure the systems success. In syllable-based systems, postprocessing is highly effective to increase the accuracy of the systems. In testing operation made with 200 Turkish words, the best accuracy rate of syllable-based systems are 94.2%, 88%, 82.6% and 90.8% for DTW, MLP, HMM and SVM respectively. But, in word-based systems, the accuracy rates for DTW, MLP, HMM and SVM are measured as 96%, 82.6%, 89.4% and 90.7% respectively.