Writing
Research notes, methodology discussions, and updates from the CogniHuman team.
Voice AI systems are evaluated almost exclusively on English benchmarks. LibriSpeech, CommonVoice, VCTK — excellent datasets, but they measure performance in a language spoken by a fraction of the world. India has 22 scheduled languages. Hundreds of millions of people speak dialects that have never been included in a standardized evaluation framework. BashaEval is our attempt to change that.
Most Voice AI research assumes cloud infrastructure. An AWS instance with an A100 GPU. A Google Cloud endpoint. A managed API. We think this is the wrong starting point for research that aims to serve communities with limited connectivity. Here is why we are building for CPUs first.
A short note on why a group of researchers registered a Section 8 not-for-profit foundation and what we believe is genuinely missing from the open Voice AI ecosystem in India — and what we plan to do about it.