What we learnt building a voice agent on two frontier Indian AI models

It took The Ken three days, a high-end laptop computer, and little cash to construct a voice-transcription AI agent on India’s state-funded massive language mannequin BharatGen. 

BharatGen is considered one of 12 firms that the federal government is backing to develop a “sovereign” AI ecosystem. However it’s considered one of solely two chasing frontier capabilities. The opposite is Sarvam. 

The Ken examined each, coding over a weekend with two programmers to construct an agent for numerous Indic languages and evaluating them with a worldwide benchmark, Scribe v2 from the US-based Eleven Labs. 

BharatGen took eternally to obtain from Hugging FaceHugging FaceA web site the place AI firms and researchers publish their fashions for anybody to obtain and use, and several other makes an attempt to put in. When the agent constructed on it did work, it interpreted background noises as Kannada or Telugu or Urdu, repeatedly glitched in transcribing a Hinglish interview, and took double the time than its friends to supply any end result. 

Creating the identical agent on Sarvam took a day, an API key, a couple of traces of code, and about Rs 45 per hour. Eleven Labs had a fair quicker course of at lower than half the fee. And the outcomes for each had been way more correct.

The hole is starker exterior check environments. In March, for instance, Maharashtra signed agreements with each Sarvam and BharatGen to deploy them in authorities capabilities. By August, Sarvam had rolled out Indus, a ChatGPT-style assistant for presidency and enterprise use, to 2,500 authorities officers in 22 languages. BharatGen’s AI assistant MahaGPT stays at pilot stage. In the meantime, on 29 August, Karnataka’s IT minister met with Eleven Labs executives to discover the usage of its voice AI throughout skilling, healthcare, and citizen providers. 

The engineers who helped The Ken construct the AI agent stated they might all the time desire utilizing Eleven Labs—or Sarvam for rarer languages—over BharatGen, though the latter was free. 

“The downloading, the debugging,” they defined, “it’s simply not value it proper now.” 

That’s not essentially a verdict on high quality, however an impression of how the opponents are structured.