Learnings and innovations in speech recognition

Abstract

In the last ten years, speech recognition has evolved from a science fiction dream to a widespread input method for mobile devices. In this talk, I will describe how speech recognition works, the problems we have solved and the challenges that remain. I will touch upon some of Google's main efforts in language and pronunciation modeling, and describe how the adoption of neural networks for acoustic modeling marked the beginning of a technology revolution in the field, with approaches such as Long Short Term Memory models and Connectionist Temporal Classification. I will also share my learnings on how Machine Learning and Human Knowledge can be harmoniously combined to build state-of-the-art technology that helps and delights users across the world.