Speech Recognition Using Novel Diatonic Frequency Cepstral Coefficients and Hybrid Neuro Fuzzy Classifier

Speech recognition is the ability of the machine to identify spoken words and classify them into appropriate category. First stage in the process of speech recognition is the extraction of appropriate features from the recorded words. We propose a novel algorithm for feature extraction using diatonic frequency cepstral coefficients. Diatonic frequencies are derived from a musical scale called as diatonic scale. The scale is based on harmonics of sound and models nonlinear behavior of human auditory filter. After feature extraction, the next classification stage uses a hybrid classifier using artificial neural network and fuzzy logic. If the difference between prediction values available at the output of the neural network is less, the classifier matches wrong patterns. Proposed algorithm overcomes this drawback using fuzzy logic. Proposed hybrid classifier improves the recognition rate significantly over existing classifiers. Test bed used in the experimentation focuses on Marathi language. It is the native language spoken in the state of Maharashtra.