Mike Cohen, who leads Google's speech technology efforts, and Franz Och, machine translation researcher, chat with Alfred Spector, VP of Research and Special Initiatives at Google, about two technologies that might seem unrelated to Google's core competency. Both statistical machine translation and speech recognition are search problems and Google's computer infrastructure can process large amounts of data that are needed to build language models. Another big advantage for Google is that it has popular services that generate a lot of useful data.
"When we first created GOOG-411, we had no speech data. Because we had so much query data here at Google (textual queries that people had typed to Google Maps), we could already train a pretty good language model. Now, obviously, text is a little different than speech and now that we've also trained on speech, we have better performance than we had back then, but even out of the box we could get good performance on that problem because we had so much textual data," says Mike Cohen.