Here's what I'm going to do to take a deeper look into the prosodic nature of human-computer dialogues:

• previous work:
  • As an undergrad I annotated a portion of the Communicator 2000 corpus and studied user behavior wrt errorful regions of dialog (ICSLP 2002).
  • As far as prosody is concerned, at ASRU 2003 I looked at acoustic features global to an utterance: a variety of pitch, rms, and spectral features that can be seen as acoustic correlates of prosody.
• Now I want to look at smaller components with in an utterance
  • pitch features of pitch contours of voiced regions
    • f0 min, max, range, median, and average for each voiced region
    • f0 slope, rise portion slope, fall portion slope, magnitude of slope (rate of change)
    • pitch contour characterizations: look into ToBi and Fujisaki model.
  • energy features
    • changes in energy between voiced segments
    • using energy from band-limited filters to calculate speech rate (alternative: use spectral moment).
  • spectral features
    • using spectral moment to calc speech rate.
    • changes to a given phonetic context given a prosodic or emotional context (like Chul and Serdar's work)
• Tools
• Here's what I have to do:
  • Get file list for all the COM2000 utterances.
  • Make feature files
    • Global features:
      • similar to what I did before.
      • incorporate tags in the data structure (maybe some other stuff too, eg maxima, minima, and/or count/freq of the local features w/in an utterance).
    • Local features: (main focus of this study)
      • divide utterance into voiced regions/unvoiced regions (or voiced/unvoiced/silence)
        • time indeces, duration.
        • min, max, range f0
        • slope measurements (avg slope, rise/fall slope, rate of change)
        • characterize pitch contour
    • Arrangement of the feature files?
      • list or hash
    • Management of data
      • need to query and sort efficiently DB or program or XML metadata.