SEGMENTATION AND RECOGNITION OF CONTINUOUS SPEECH

          Identification of the boundaries between different sound units in a speech signal is called segmentation. The first task in the development of any voice activation system from the scratch, is segmentation. Today this process is done manually in most speech systems.

      We propose a new algorithm to automatically determine segment boundaries in a speech signal. This algorithm is based on the use of the minimum phase group delay function. The speech signal is processed to derive the short term energy function. Root cepstral based group delay functions are derived for this energy function. The group delay function is further processed to derive the segment boundaries.

      Connected Digit and continuous speech recognition systems are being implemented in which the speech signal is segmented at digit and syllable boundaries respectively and an HMM based isolated style recognition system is used for recognition of the segments.