Continuous speech recognition using automatically segmented data at syllabic units

We propose an alternative approach for continuous speech recognition where the segmentation and recognition tasks are separated. Syllable is considered as a unit for both segmentation and recognition. Using minimum phase group delay function based approach, the speech signal is segmented at boundaries of syllabic units and a syllable based isolated style HMM recognition system has been implemented for two Indian languages. To address the errors in recognition due to shift in segment boundaries and merger of syllabic units, Viterbi algorithm based approaches are proposed.