ISCA Archive Interspeech 2015
ISCA Archive Interspeech 2015

Handling derivative filterbank features in bounded-marginalization-based missing data automatic speech recognition

Marco Kühne

This paper extends the familiar missing-data bounded-marginalization technique from static to dynamic filterbank features for noise robust automatic speech recognition. Based on a well-known theorem from Statistics it is shown how the reliability of derivative filterbank features can be expressed in form of a probability density function. As another contribution, the corresponding HMM state emission likelihood equation (bounded-marginalization rule) for dynamic features is derived in closed-form. On the CHiME corpus the new approach showed a superior accuracy compared to previously proposed heuristics for handling missing dynamic features. To the author's best knowledge, the achieved average accuracy of 92.58% is the best result so far reported for the 2011 CHiME Challenge task.