Maya Research releases Apache 2.0 Maya 2 TTS covering eleven Indian languages

What shipped

Maya Research, an Indian startup that has raised $1.9 million in seed funding according to the dev.to post covering the release, published open weights for its Maya 2 Native text-to-speech model on 25 October 2026. The model runs 2.5 billion parameters, weighs roughly 5 GB in BF16, generates 24 kHz mono audio in 80 ms streaming frames, and is licensed under Apache 2.0, which permits commercial use and fine-tuning.

Twelve languages are supported: English plus eleven Indic languages, namely Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu. The release includes twenty built-in voices and zero-shot voice cloning from a reference clip of three to ten seconds.

One unexpected capability is cross-lingual cloning: as the dev.to write-up highlights, a Hindi voice sample can be used to synthesise Tamil speech in the same speaker's voice. The voice carries over rather than the content being translated. Whether this holds up reliably in production remains untested.

The benchmark picture

Maya's CEO, Dheemanth Reddy, claimed on X that Maya 2 outperforms ElevenLabs v3, Cartesia Sonic 3.5 and Sarvam's Bulbul v3 across Indian languages. As the dev.to author points out, that comparison is effectively a press release, since the post did not show how the test was run.

Third-party numbers tell a narrower story. VoiceArena's Hindi leaderboard places Maya 2 Native second with an Elo of 1047, behind Cartesia Sonic 3.6 at 1094 and ahead of Bulbul v3 at 957. On Artificial Analysis, which tracks 103 speech models globally, Maya 2 Global sits at 1048 Elo — rank 28, below ElevenLabs v3 and much of the rest of the field.

The fair reading offered in the write-up is that Maya 2 is genuinely competitive within Indian languages, while a 2.5-billion-parameter model was never likely to top a global leaderboard.

Latency and integration details

For voice agents, latency matters as much as quality. Maya claims 75 ms time to first byte, with 7.3 ms of variance between invocations, measured from Bangalore to the company's Mumbai region. The documentation adds a detail many will miss: first audio takes 389 ms when a new connection is opened per request, versus 88 ms when the connection is reused — a roughly fourfold difference that can make a live demo feel broken.

Several implementation quirks are documented. The API returns raw 16-bit little-endian mono PCM at 24 kHz with no file header, so writing the bytes directly to a .wav file produces something most players cannot open. Voice names are case-sensitive: "Ananya" and "ananya" yield different audio, because the model interpolates whatever string it receives. Emotion tags such as [excited] work only in a separate model, Maya 2 Native Emotional, which supports ten tags. Code-mixed text, which Indian users type routinely, is handled by omitting the language field and relying on auto-detection.

There is also a naming leftover: the Hugging Face card states that the package, Python class and configuration filename still use the maya3 name (Maya3TTS and maya3_config.json) even though the release is called Maya 2, suggesting a version three that was renamed or skipped.

How open is it in practice

A 5 GB BF16 checkpoint cannot realistically run on a laptop with 8 GB of RAM and no GPU, as the dev.to author notes. The licence is generous, but for most developers the hosted API remains the practical access route. The weights matter mainly to research labs and startups with the hardware to fine-tune for a specific dialect.

Why it matters

India has roughly 1.5 billion people and 22 official languages, and typing most of those languages on a phone is slow enough that many users fall back to English mid-conversation. In that context, voice is not an optional accessibility layer; it is often the input method that does not disadvantage speakers of Indic languages. An Apache 2.0 model covering eleven of them is something few comparably funded labs anywhere have shipped for non-English languages. The headline benchmark claim still needs independent replication, and the checkpoint size limits who can run it locally, but for developers building Hindi, Tamil or Marathi voice agents a competitive open alternative meaningfully changes the economics.