Speech Group Projects

In general, this page is always "under construction". However, current Speech Group projects include the following:

• Effective Affordable Reusable Speech (EARS): This is the major DARPA speech program that started in 2002. We are participating in both major components of this program, namely:
  • Rich Transcription of Conversational Speech - in collaboration with SRI and the University of Washington, we are working to generate readable transcriptions of conversational speech in multiple languages. "Readable" here means incorporating capitalization, punctuation, and speaker markers; but it also means making major improvements in speech recognition performance, since word errors are still significant in this type of task.
  • Novel Approaches - we are studying both replacements of the standard spectral envelope as the speech representation of choice (typically with cepstral transformation). This includes work on the acoustic "front end", but also includes research on statistical modeling for the new features that are being generated.

• The Meeting Recorder project, which seeks to develop speech recognizers that would be useful in conventional meeting contexts, as well as information retrieval and other applications that such recognition would make possible. One component of this work is the SpeechCorder, which is a project to design a portable digital tape recorder that uses robust speech recognition to create an indexable and annotable word stream for archiving meetings. SpeechCorder is also associated with the IRAM project.
• Speaker Recognition:  The goal of this project is to use higher-level features (such as word usage, prosodic characteristics, pronunciation patterns, idiosyncratic laughs or other non-speech events) to improve speaker recogniton.
  


• In addition, there are a number of ongoing projects (and others just starting up) in a range of areas pertaining to speech processing that is robust to noise and reverberation, to the automatic derivation of speech categories, to the incorporation of prosodic features for predicting punctuation, disfluencies, and overlaps, and the automatic classification of frustration during human-machine interaction.
And see here for a number of earlier projects.