Introduction

On October 14, Maya Research (India) put Maya-2 in front of builders: a closed, real-time streaming text-to-speech stack aimed at Indian languages, with free open weights promised “very soon.” The CEO’s launch post on X frames it as fully open-weight Indic voice AI that beats ElevenLabs v3, Cartesia Sonic 3.5, and Bulbul v3 on Indian speech. Treat that leaderboard claim as vendor-reported until you hear independent tests — what you can use today is the hosted API, Maya Studio, and Baseten dedicated deploys.

This is not Sesame’s “Maya” voice, and it is not a rehash of Maya Research’s earlier open Maya 1 / Veena cards. Maya-2 is the new flagship family. Weights are not on Hugging Face yet (maya-research still lists Maya 1 and Veena). Baseten lists the license as Proprietary – Maya Research.

What Shipped

  • Family: Maya 2 Native (flagship, #1 real-time Hindi / #2 overall Elo 1086 on Voice Arena), Maya 2 Global (#3 real-time English / #4 overall Elo 1024), and Maya 2 Flash (speed-first live conversation cut).
  • Size: 2B parameters (Baseten model card).
  • Languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, and Indian English — 11 total on the site and Baseten card.
  • Audio path: 24 kHz PCM streamed; company latency claim of 75 ms time to first byte (Bangalore → Mumbai region), with 7.3 ms call-to-call variation.
  • Request caps (Baseten): up to 130 words of text in, up to 100 seconds of audio out per request.
  • Runnable today: POST https://tts.mayaresearch.ai/v1/tts and WebSocket /v1/tts/stream (API keys via Maya / email), Maya Studio, plus US and India dedicated Baseten deploys. Regions called out on the site: Mumbai and Iowa.

curl -X POST https://tts.mayaresearch.ai/v1/tts \
  -H "Authorization: Bearer $MAYA_API_KEY" \
  -H "content-type: application/json" \
  -d '{"voice":"Ananya",
       "text":"नमस्ते! आपका ऑर्डर कल पहुँच जाएगा।",
       "language":"hi"}' \
  --output out.pcm

Open Weights: Announced, Not Live

The launch thread leads with “fully open-weight” and “completely free for every builder very soon.” That promise matters — Maya 1 is already Apache 2.0 and Veena is still open — but Maya-2 weights are not downloadable yet. Until a Hugging Face card or GitHub release appears, ship this as closed API today, open weights pending. A follow-up “growth milestone… share tomorrow” post from the same founder is worth watching for the actual weight drop.

How It Fits the Indian TTS Lane

ArtRealmAI already covered Sarvam’s Bulbul V4 Flash (sub-250 ms production TTFA, API-only). Maya-2 is competing in the same voice-agent lane with a stronger Voice Arena Hindi story and a public open-weight tease. If you need India-region latency and native-language rhythm for IVR, BFSI, ecommerce, or agent stacks, try Maya’s Mumbai region and A/B against Bulbul Flash on the same Hinglish and code-switched scripts. If you need weights on disk today, stay on Maya 1 / Veena or another open TTS until Maya-2 lands on HF.

Conclusion

Maya-2 is a same-day runnable Indian-language streaming TTS drop with a clear Arena ranking story and a honest open-weights asterisk. Hit the API or Baseten with a short Hindi or Tamil order-status line, listen for accent and number handling, and keep an eye on maya-research for the free weights the launch promised.

Original source:

Site: mayaresearch.ai · Baseten: Maya 2

—Anabel ♡