Identification of vowel-non vowel letter with artificial neural network and sound-text synchronization at syllable level

In the context of this thesis, it was aimed to synchronize text-voice in syllable-size by discriminating vowels from non-vowel in a word. Within the scope of this developed method, two different artificial neural network models were used as classifier, and Mel-frequency cepstrum coefficients were chosen for voice feature extraction. Pattern recognition network and Layer-recurrent network were used for this process.In this speech recognition system, which is created as Speaker-depended, vowel and consonant samples were taken at the first stage, and the neural networks were trained by features extracting of these samples. In the second stage, the network, which gave the best performance, was determined by testing. While the layer-recurrent neural network achieved an average of %97.43 success in the classification of vowel-consonant, the pattern recognition network achieved %98.82 in identification of vowels, %97.27 in consonants and %98.04 on average. In the third stage, in which the network showing the best performance was chosen, the sounds consisting of 10-word vectors were supplied separately to the network and the vowels in the words were successfully detected. In the fourth step, these words are separated into syllables with the developed syllable algorithm in which created in an interface program, the text and voice files that corresponding to these words are simultaneously emphasized in the syllable size. As a result, pattern recognition networks show a successful performance in the recognition of vowels, and it was also shown that text-voice synchronization can be done by detecting vowels in the text.