Teaching a machine to hear Hindi
Amazon Alexa speech recognition
Speech-recognition training data has a requirement most document work does not: it has to match how people actually talk. Written Hindi and spoken Hindi diverge sharply in word order, register and borrowed vocabulary, and a corpus built from formal prose teaches the model the wrong language.
Delivered alongside Google product translation covering interface strings, legal terms and marketing copy in both Hindi and Urdu, and product content for Facebook.