Your Speech Model Isn't Bad - It's Just Built for Someone Else

​Speech recognition often fails because it doesn't consider the diversity of real voices that arises from differences in age, gender, dialect, language and many other factors. Your own voice changes as you grow, or sometimes even when you change countries! In this talk, you'll discover why having a deep understanding of your end-user and building accordingly is crucial. We will explore this through my personal experience of building speech-to-text models in Gujarati and Hindi for children (primarily rural) in Gujarat and Rajasthan respectively.

Speaker: Ayush Deva
LinkedIn: /ayushdeva
X: @AyushDeva
Current: ML Engineer @ Google
Previously: ML Engineer @ Wadhwani AI, B.Tech - IIIT Hyderabad