Including prosodic cues in ASR systems

Several aspects related to production, as well as natural perception of speech, have gradually been incorporated to automatic speech recognition systems. Nevertheless, the set of speech prosodic characteristics has not been used for the time being in an explicit way in the recognition process itself. In this work, an analysis of the prosody’s three most important parameters: energy, fundamental frequency and duration, is presented with a method to incorporate this information into automatic speech recognition. Prosodic-accentual features are incorporated in a hidden Markov models recognizer. Their theoretical formulation and experimental setup are presented. Several experiments are developed to show the method behavior in a Spanish continuous speech database. From this understanding and with other database subsets, the overall results provide a word recognition error reduction that would reach more than 30% when prosodic-accentual cues are incorporated.