Sixth International Conference on Spoken Language Processing
We propose a new Bayesian approach to cope with a class of difficult robust speech recognition problems for those applications which only involve a couple of utterances, but every utterance involves a distinct yet complicated "distortion channel" from the intended message a speaker wants to convey to the received signal of a speech recognizer. It works by composing on-the-fly a testing-condition dependent prior distribution of HMM parameters for each testing utterance from a set of prior distributions elicited off-line from a rich set of training data. Encouraging results are obtained by applying this new method to a task of speaker independent continuous Mandarin speech recognition for speakers with different degree of accents.
Bibliographic reference. Huo, Qiang / Ma, Bin (2000): "Robust speech recognition based on off-line elicitation of multiple priors and on-line adaptive prior fusion", In ICSLP-2000, vol.4, 480-483.