Sunil Maurya

Sunil MauryaOct 29, 2026

In many ways, AI4Bhārat is the open-source backbone of India's Voice AI. One Indian Institute of Technology, Madras lab, launched in 2019, built much of the free STT, TTS and LLM stack Indian builders use today. Its IndicVoices dataset alone has 12,000 hours of real speech. One of its co-founders went on to start Sarvam, now valued at $1.5B. I went through its people, models and datasets one by one. Here are the highlights: 1/ Sarvam - Pratyush Kumar & Vivek Raghavan Voice and language AI for India, valued at $1.5B. Kumar co-founded AI4Bharat in 2019, then started Sarvam with Raghavan. 2/ Bodhan AI - Mitesh Khapra A Ministry of Education Centre of Excellence at IIT Madras, with Indic-Transcribe, Indic-Speak, Indic-Translate and Indic-OCR. Khapra heads AI4Bharat and was named to a TIME 100 list. 3/ IndicVoices - Tahir Javed 12,000 hours of real Indian speech, the dataset behind much of the open Indic STT work. He also worked on IndicWav2Vec. 4/ Vistaar - Kaushal Bhogale & Abhigyan Raman An open benchmark for Indian-language STT, alongside the IndicWhisper models. Raman also worked on IndicSUPERB. 5/ IndicWav2Vec - Sumanth Doddapaneni & Gowtham Ramesh Pretrained speech models for Indian languages. 6/ Indic TTS - Ashwin Sankar Open text-to-speech models for Indian languages, built on VITS. 7/ Airavata - Jay Gala, Thanmay Jayakumar, Raj Dabre & Rudra Murthy An open instruction-tuned Hindi LLM. 8/ IndicNLPSuite - Divyanshu Kakwani Open text corpora and models for Indian languages. 9/ IndicTrans2 - Anoop Kunchukuttan AI4Bharat co-founder, now at Microsoft. Leads open translation across Indian languages. 10/ The funders - Nandan Nilekani Backed through Nilekani Philanthropies and EkStep, as part of about ₹70 Cr in funding. Bhashini, Microsoft, Google and NVIDIA have also backed the lab's work. Also on the shelf: BhasaAnuvaad: 44,400 hours Shrutilipi: 6,400 hours Nirantar: 3,240 hours Kathbath: 1,684 hours The lab also runs Voice of India with Josh Talks, a public STT leaderboard: 14 systems, 675 districts, 36,691 speakers, 536 hours of speech. Sarvam Audio has the lowest error rate in 13 of 15 languages. Sarvam was started by an AI4Bharat co-founder, so read the methodology yourself. Who else came out of AI4Bhārat? PS: I am building Unio - The Voice AI Community, it's a WhatsApp group, invite only. Want to get added? DM me.

Sunil MauryaOct 22, 2026

A model that is not an STT, TTS or LLM is being tested by builders in our Voice AI community. It's called Jev, by TypeSafe AI, a newly launched decision model. It returns a decision with a confidence score instead of generated text, and it's said to cut cost and make decisions very fast. I'm a non techie, so I had more questions than answers. Can a decision model replace an LLM for simple customer replies? The engineers told me usually no. You probably don't need an LLM to hear the word "yes", because rules and small classifiers win on plain yes or no. They also told me even "yes" isn't always simple, because it depends on what the agent just asked. And a wrong yes on a policy renewal costs more than the inference you saved. Could it improve turn detection? Wait too long and the agent feels slow, cut in early and it interrupts the customer. Builders in our community pointed out that sending audio to a remote model can cost more time than it saves. A smaller on device model may fit better, and no one in our group has shared results yet. Could it help with complex workflow routing or policy decisions? This is where builders felt a decision model is worth testing, because choosing between workflows, tools and policies is a real decision problem. Then came the part that changed my view. One team in our community reported using Jev as a judge across 34 voice agent metrics. After up to three rounds of refinement, 27 reached 96% or higher on calls the judge had never seen, and 7 stayed stubborn. Their reported numbers, not independently verified. But it made me think judging calls may be a better use than replacing the LLM during them. A model doesn't belong in your pipeline because it's smaller. It belongs there if latency, accuracy and cost per successful call all improve. If you've implemented Jev, share what you built. If you've tested something similar, tell me what you found. 👇 PS: I am building Unio - The Voice AI Community, it's a WhatsApp group, invite only. Want to get added? DM me.

Sunil MauryaOct 6, 2026

gnani.ai does 1.5 billion minutes a month. Number two does about 60 million. That is the gap at the top of Indian Voice AI. We just updated the Top 10 Voice AI Companies in India, across industries, ranked by estimated monthly minutes in production. gnani.ai, 1.5B minutes SquadStack.ai, ~60.9M minutes Blue Machines AI, ~56.9M minutes CredResolve, ~36.8M minutes IGNOSIS AI, ~25.6M minutes Bolna (YC F25), ~20.5M minutes YuVerse (Yubi Group), ~17.1M minutes Ringg AI, ~14.9M minutes Fundamento, ~14.3 minutes Vocallabs AI, ~11.4M minutes Every volume was confirmed directly by each founder. Here is what stands out to me. Gnani.AI alone is about 85% of the volume in this top 10. The market is concentrated at the very top. Between number two and number ten, the gap is only about 6x. That is where the real race is. Not all calls are equal. SquadStack reports 107M calls, but a dialed call nobody picks up is not a conversation. So we convert everything to estimated minutes to compare on one scale. That is why CredResolve, with 21M connected calls, ranks above Ignosis with 45M dialed calls. Volume gets you on the list, but trust keeps you there. The teams that can prove their numbers will win the next round of buyers. The ranks are an indicative Unio estimate. The full list and method are here: https://lnkd.in/dFAUq6Uk PS: I am building Unio - The Voice AI Community, it's a WhatsApp group, invite only. Want to get added? DM me.