Sixth European Conference on Speech Communication and Technology
We continue our study of the use of fabricated data in the investigation of speech recognition algorithms. After reviewing the basic data generation algorithm and some earlier results involving the recognition of fabricated conversational speech data, we go on to describe some new and intriguing experiments concerning on the one hand, training acoustic models from simulated speech data, and on the other, recognizing fabricated data with different amounts of context in the language model. Among other things, we conclude that standard training algorithms are remarkably good at recovering the underlying structure in acoustic models when they are given 30 hours of data generated from those models. We also propose an alternative to erplexity for measuring the quality of a language model - the word error rate on fabricated data - that takes into account the inherent acoustic confusability of words.
Full Paper (PDF) Gnu-Zipped Postscript
Bibliographic reference. McAllaster, Don / Gillick, Larry (1999): "Studies in acoustic training and language modeling using simulated speech data", In EUROSPEECH'99, 1787-1790.