Maya (Maya Research)
Open-weight expressive voice AI from India — Maya-1 in English, hosted Maya 2 in eleven Indian languages.
Overview
Maya Research, a Bengaluru start-up, makes Maya-1, a 3B-parameter open-source TTS model on a Llama-style transformer with inline emotion tags (17 in its official list) and natural-language voice design, released Apache 2.0 in Nov 2025 and runnable on a single 16GB GPU; its open weights are English-only. Its hosted Maya 2 models speak eleven Indian languages.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 28 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Full model access
- Commercial use allowed
- English with multi-accent support
- Keys from mayaresearch.ai/studio/developers
- Two voices, 24 kHz streaming
Use cases
What you can produce with Maya (Maya Research)
- Emotionally expressive TTS audio at 24 kHz, with sub-100 ms latency claimed by the vendor
- Inline emotion-tagged outputs — 17 tags in the official list, including laugh and cry
- Zero-shot voice generation from natural-language persona descriptions (age, accent, tone, character)
- Self-hosted inference with vLLM, plus community-built GGUF files and ComfyUI nodes
- Open-weight Apache 2.0 model weights downloadable from Hugging Face (maya-research/maya1)
- Hindi, Tamil, Telugu and eight more Indian languages through the hosted Maya 2 API
- API keys from mayaresearch.ai/studio/developers once access is granted
ASEAN Perspective
Maya (Maya Research) in Southeast Asia
Maya Research explicitly targets Southeast Asia as a key market alongside India and MENA, and the consumer app has already crossed 3 million installs across these three regions combined. The company's mission — building voice AI for the 90% of the world that mainstream English-first systems ignore — maps directly onto ASEAN's linguistic diversity, where hundreds of regional languages and accents are underserved. Current open-source weights cover English; Hindi and Telugu are the next confirmed languages, with broader Indic and Asian language support planned for mid-2026. ASEAN developers can self-host Maya-1 today for English-language voice apps at zero per-call cost, making it a strong candidate for regional accessibility and low-bandwidth use cases.
Maya-1 is a genuinely impressive technical feat from a small Indian startup that trained a competitive open-weight TTS model on cloud credits. On the live Artificial Analysis leaderboard in September 2026 it sits seventh of 16 open-weight models at a Quality Elo of 1,046 — reported #2 at launch, but the field has crowded in since. Its inline emotion tags and zero-shot natural-language voice design still stand out. For self-hosters, Apache 2.0 means no per-character fees and no vendor lock-in, which matters in cost-sensitive ASEAN deployments. Caveats: the open weights cover English with multi-accent support only — the Indian languages live in the hosted Maya 2 API, which has no public price and currently needs a promo code — and self-hosting needs a 16GB-VRAM GPU.
What we found
A small Bengaluru team built a text-to-speech model on cloud credits that briefly out-ranked entries from labs with nine-figure budgets. Maya-1, released under Apache 2.0 in November 2025, was reported at #2 among open-weight models on the Artificial Analysis Speech Arena at launch; we could not confirm that launch-day table, which the web archive does not capture.
That rank hasn't held: on 28 September 2026 the live leaderboard put Maya-1 seventh of 16 open-weight models, at a Quality Elo of 1,046. The field simply got more crowded, which is a compliment to the category Maya helped kick off as much as a knock on the model. What hasn't changed is the pitch — inline emotion tags, 17 in the official list, zero-shot voice design from a plain-language description like "warm 40-year-old male, slight Southern Indian accent," and, by the vendor's own account, sub-100ms latency on a single 16GB GPU.
Traction is real for a company this young: Hugging Face counts about 354,000 all-time downloads of Maya-1 alone, and South Park Commons led a $1.9M seed round; the company also claims a consumer app past 3 million installs, but names no app we could check. The Maya-1 open weights remain English-only with multi-accent support. The Indian languages went into the hosted product instead: Maya 2 streams speech in eleven Indian languages, Hindi, Tamil and Telugu among them, in two voices, and its API has no public price — access currently needs a promo code. Self-hosting Maya-1 needs a 16GB-VRAM GPU, which rules out casual tinkering on a laptop.
Best read as a genuinely strong open-weight option for cost-sensitive ASEAN builders today, with the hosted Maya 2 line the thing to watch through the rest of 2026.
Maya says Maya 2 took the top Hindi spot on Voice Arena on more than 15,000 human votes, ahead of ElevenLabs and Cartesia. It runs on Baseten infrastructure in the US and India, and the vendor measures first audio at 83 milliseconds over WebSocket.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Maya Research built the entire Maya-1 model without venture capital, using only free cloud credits from AWS, Google, and Microsoft Azure.
- Both co-founders are 23-year-old NYU graduates from rural Andhra Pradesh — Dheemanth Reddy from Vempalli and Bharath Kumar Kakumani from Ongole.
- Maya Research is the only Indian company listed on the Artificial Analysis Speech Arena leaderboard as of 2025.
- Maya-1 supports over 20 inline emotion tags that physically alter the audio waveform — changing pitch, breath timing, and pacing — rather than just applying a filter on top.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Maya (Maya Research)'s official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Maya (Maya Research) unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Maya (Maya Research) directly →
Spotted something out of date? Suggest an update →
Related reading
More in LLMs & Chat
