Thursday, April 23, 15:30 - 18:00, Location: Show and Tell Area A
Yun-Cheng Ju
Replying text messages on mobile devices, especially while driving, is a big challenge for obvious reasons. Even with the help of speech recognition, there is no practical yet safe correction UI to recover speech recognition mistakes. We tackle this problem by using voice search techniques: (1) we built an efficient SLM trained from a collection of SMS messages to improve robustness; and (2) we use TF/IDF to map recognition result to the closest one(s) from thousands of popular messages we have collected; and (3) we present the high quality, relevant, scrubbed template to eliminate the need for correcting (very likely) misrecognized words.
A prototype of this system is developed on a Windows XP machine (notebook) where both speech recognition and spoken language understanding are performed locally on the device. It is part of a ongoing interactive in-car infotainment system project (Commute-UX) where drivers can use natural speech and/or multimodal user interface to reply/delete text messages or call back the sender.