|RESPITE:Events : Meeting, Sep 2000:Presentations: Dan Ellis|
We have continued to investigate the 'tandem' architecture, where
the outputs of a neural net trained to produce context-independent
phone posterior probabilities are used as features for a conventional
GMM/HMM recognizer. This approach continues to be successful for the
'mismatched' Aurora-2000 conditions, and preliminary results from the
relatively large-vocabulary DARPA SPINE task are promising.
Other multistream work at ICSI includes investigations into using mutual information as a guide for choosing streams and combination methods, using multiple feature streams independently optimized for different noise conditions, and more investigations of ways of combining a 'full combination' array of multiband nets to approach oracle results.
I will also give a brief overview of the ICSI initiative in Meeting Recording: we have been collecting data from real meetings with a view to designing recognizers for this scenario for a variety of applications.