An Indian start-up wants to make voice AI sound native in Indian languages. Dheemanth Reddy, co-founder and CEO of Maya Research, announced Maya-2, which he calls the world’s best Indic voice AI model and says is fully open-weight.
The company says it beats rivals, but the evidence so far comes mostly from the company and a profile article.
More on this topic: TechnoSports Maya Research coverage.
Maya-2: Key Details
| Claim | Detail |
|---|---|
| Model | Maya-2 |
| Company | Maya Research |
| Compared against | ElevenLabs v3, Cartesia Sonic 3.5, Bulbul v3 |
| Languages | Indian languages, 22 official languages cited |
| Licence | Open-weight |
| Price | Free for builders “very soon” |
| Voice Arena (Hindi) | Ranked first after 15,000+ blind votes, per a profile |
What Was Announced
Reddy posted that the model officially outperforms ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3 across Indian languages, and said it will be completely free for every builder very soon. He framed the future of AI for India as voice-first, with 22 official languages.
An Analytics Insight profile says Maya 2 ranked first for Hindi on Voice Arena after more than 15,000 blind human votes. The same article notes it does not independently verify those rankings.
Why Indic Voice Matters
Most voice models are tuned for English first. Indian users switch between languages mid-sentence and use many accents, which makes native quality hard. Open-weight models let start-ups build call-centre bots, tutors and accessibility tools without paying per character.
Maya Research is backed by South Park Commons, and Reddy has received an Emergent Ventures grant, according to the same profile.
How to Judge the Claim
Blind listening tests are a fair way to compare voices, but results depend on the sample of listeners and the prompts. Try the model on your own scripts and languages, and listen for pronunciation, emotion and code-switching.
Read about speech synthesis if you want the basics, and see our voice AI coverage.
What Builders Should Check
Open-weight does not always mean free to use commercially. Read the licence for limits on commercial use, voice cloning and redistribution before building a product.
Also test latency, since real-time voice agents need fast first-audio response, and check how well the model handles numbers, names and mixed-language text.
What Is Not Confirmed
- Independent tests confirming it beats ElevenLabs v3, Cartesia Sonic 3.5 and Bulbul v3.
- The licence terms and release date of the weights.
- Which of the 22 languages are fully supported.
- The “free very soon” timeline.
Quick Answers
Who built Maya-2?
Maya Research, co-founded by Dheemanth Reddy.
Is Maya-2 free?
The company says it will be free for builders very soon.
Is the ranking independent?
The Voice Arena result comes from public votes, but the profile article did not verify it.
Source: Dheemanth Reddy on X, Analytics Insight, Speech synthesis on Wikipedia.
How Builders Can Test Voice Models
Run the same set of scripts through each model you consider, including names, numbers, English words in the middle of Hindi sentences and emotional lines. Ask native speakers to rate them blind, as the Voice Arena does, rather than trusting a single demo.
Also measure time to first audio and cost per hour of speech. A model that sounds slightly better but responds slowly may be wrong for live phone agents, while it may be perfect for audiobooks or dubbing.
Bottom Line
Maya-2 could be a real step for Indian voice AI if the claims hold in independent tests. Right now, most of the evidence is from the company and a profile piece.
Try the model when the weights are public and judge it on your own languages and use cases.
