Solving Resource-Constrained Indic Language Understanding Problem Using Unsupervised Learning
, Airtel.com
, Airtel.com
In India, 135 Cr people speak 13 languages with several different dialects. At Airtel, we're building voicebots that understand, infer, and speak these languages. Typically, state of the art voicebots need 10,000 hours of annotated speech-to-text audio data to train, but Indic languages don’t have such high quantities of annotated data available, posing the challenges of getting state of the art models with low resources and fastening the training cycle. Using NVIDIA GPUs and a distributed high performance compute architecture, we're solving both challenges and building bots that understand and speak multiple languages.