Finding mel frequency cepstral coefficients, linear predictive coding methods and Turkish speech recognition using artificial neural networks and comparision of the methods

In this thesis, speaker dependent, isolated words automatic speech recognition systems are developed using MATLAB program. Speech recognition systems consist of feature extraction and classification main parts. In this study, Mel Frequency Cepstral Coefficients and Linear Predictive Coding techniques are used as feature extraction methods and Dynamic Time Warping and Artificial Neural Networks are used as classification techniques. Mel Frequency Cepstral Coefficients and Linear Predictive Coding methods and algorithms are used to find feature coefficients in classification and speech recognition methods using Dynamic Time Warping and Artificial Neural Network. Therefore, 4 different speech recognition systems are developed. The output responses of these 4 methods are compared. It is observed that Mel Frequency Cepstral Coefficients method gives better results than Linear Predictive Coding method among feature extraction methods. It is also observed that Dynamic Time Warping method produces better results both in terms of Word Recognition Rate and vocabulary compared to Artificial Neural Networks.