Linear predictor coefficient, power spectral analysis and two-layer feed forward network for bangla speech recognition

Within this research we investigates the tools and techniques how correctly recognize Bangla (Bengali) language. For feature extraction the power spectral analysis and linear predictor coefficient method considered. Speech samples taken from bangla phoneme, word, command and sentence. Each frame of a speech signal's spectrum segmented into number of fragments to take the mean absolute value from each fragment for pattern classification. A two-layer feed forward neural network with maximum likelihood method used. Features extraction by power spectral analysis gives little more accuracy for bangla phoneme and word recognition comparing to linear predictor coefficient analysis. Both methods provide average result for bangla command and sentence recognition. The mel frequency cepstral coefficient could be another approach for future research.