Computational Linguistics Colloquium

Friday, 8 November 2013, 15:15
Conference Room, Building C7.4
Note the unusual time and date!

Getting more words right than wrong in a week

Automatic speech recognition is still somewhat of a black art, and recognizers are complicated beasts that need to be laboriously trained and optimized under expert supervision. Bootstrapping a recognizer in a new language first and foremost requires language and technology experts, who laboriously optimize systems in an iterative process that requires lots of compute power, time, and data - none of which may be easy to come by. In this talk, I will describe a number of techniques that we have developed that help develop speech recognizers in new languages rapidly by leveraging information from other languages, by restructuring the acoustic model, and by optimizing features and other resources. I will also describe a framework that can learn from experts' decisions about how to develop speech recognizers, and was able to predict their choices, opening a path to an almost automatic way of building new speech recognition systems in the future.

If you would like to meet with the speaker, please contact Michael Wiegand.