Handling unexpected acoustic data

Organizer: Hynek Hermansky (IDIAP)

Panelist: Li Deng, Jeff Bilmes, Jordan Cohen, Ralf Schlueter, Herve Bourlard

Most of the "knowledge" in ASR come from training data,
(be it text or acoustic data). How should we handle unexpected acoustic 
inputs that has not been seen in the training data or not supported by a 
prior knowledge ? Such inputs do not severely impair human speech 
communication, often are of high information value, but present 
significant problems to ASR. Examples could be out-of-vocabulary, 
out-of-language, out-of-domain words, accented speech, children speech, 
accented speech, unexpected noises, e.t.c. Do we expect that gradual 
improvement of the current stochastic approach to ASR will alleviate 
this problem?