June 17, 2018

Interspeech 2018 Special Session: Low Resource Speech Recognition Challenge for Indian Languages

Baselines are built using Kaldi. Please see this README file (opens in new tab) for instructions on how to replicate the baselines and for links to lexicons for all languages.

The Word Error Rates of the baseline systems for all three languages are below:

Language GMM-HMM DNN TDNN
Tamil  33.55  25.47  19.45
Telugu  40.12  34.97  22.61
Gujarati  23.78  27.79  19.76