Maya 2 Native:
An Open Voice Model Built for the Languages India Speaks
A small Bengaluru lab just released an open, Apache-2.0 speech model for eleven Indic languages and English that ranks second for Hindi among real-time models on VoiceArena. Who built it, how it works, how it measures up, and where it still falls short.
By Serverless Ventures
Published: October 27, 2026 · Read time: ~9 minutes
The short version
- What: Maya 2 Native, an open-weights streaming text-to-speech model for 11 Indic languages and English, released October 25, 2026 under Apache-2.0.
- Who: Maya Research, a small Bengaluru lab founded in 2025 by NYU graduates Dheemanth Reddy BS and Bharath Kumar Kakumani.
- How good: second for Hindi among real-time models on VoiceArena, behind Cartesia and ahead of Sarvam, xAI and Microsoft, with a held-out character error rate of 2.86% across 6,000 test sentences.
- Why it matters: it is small enough to self-host on one 12 GB GPU, can clone a voice across languages from a few seconds of audio, and is free to use commercially.
- Watch out for: Assamese, very short turns, URLs and code-mixed text, and the lack of an audio watermark.
The bet: the next billion users will talk, not type
Dheemanth Reddy grew up in Vempalli, a small town in Andhra Pradesh, and watched his mother, educated and capable, struggle with phones that expected her to work in English. "I wanted her to be able to speak the way she already speaks and have the technology understand," he has said. "My test is still whether my mother could use what I build."
That is Maya Research's thesis in one line: the people coming online next will reach AI through speech in their own language, not through English text boxes. Reddy, who studied deep learning under Yann LeCun during a master's at NYU, leads the company as CEO; co-founder Bharath Kumar Kakumani leads the technical work as CTO. The lab was founded in 2025 in Bengaluru and is listed in South Park Commons' portfolio. It is small: the model card credits five people, the two founders plus Sai Charan Thippareddy, Prahalad Boreddy and Nikil Thalapaneni.
Maya 2 Native is its third open release. Veena (June 2025) covered Hindi and English; Maya 1 (October 2025) became, by the company's account, one of the most downloaded open speech models. Maya 2 Native widens the language set to twelve.
What it is
| Maya 2 Native | |
|---|---|
| Languages | Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu |
| Size | 2.50B parameters, about 5.0 GB of BF16 weights |
| Audio | 24 kHz mono, streamed in 80 ms frames |
| Voices | 20 built-in voices, plus zero-shot cloning from 3 to 10 seconds of reference audio, across languages |
| Interfaces | Python API, OpenAI-compatible speech endpoint, WebSocket streaming |
| License | Apache-2.0, with open weights and inference code |
| Release | 1.0, October 25, 2026 |
Two features stand out for builders. Cross-lingual cloning: a voice cloned from a few seconds of Hindi can speak Tamil. And an OpenAI-compatible speech endpoint, so an existing voice agent can swap it in without rewriting its client. Text normalization expands numbers, currency, dates and addresses into spoken form; it is on by default in the servers and needs normalize=True in Python.
How it works
Maya 2 Native treats speech as a language-modeling problem. A backbone initialized from Qwen3-1.7B-Base reads the text and a speaker embedding and predicts the first of 32 audio codebooks for each 80 ms frame. A depth decoder of about 555 million parameters fills in the other 31, and Kyutai's open Mimi codec decodes the frame into 24 kHz audio, which can be played while the next frame is generated. That is what makes it streamable. After pretraining on licensed and synthetic speech, it was tuned with direct preference optimization, ranking its own candidate outputs by how well speech recognizers understood them and how closely they matched the target voice.
The design choice is instructive. Rather than training a speech model from scratch, a five-person lab started from an open language model and an open codec, and put its effort into Indic data and preference tuning. It is a template for how small teams can compete in speech.
Go deeper: our Snack on AI deep dive, Inside Maya, India's Open Voice Model, the Small Model Is the Slow One, digs into the architecture and finds a counterintuitive bottleneck: not the 1.7B language model, but the 555M depth decoder behind it.
How it measures up
Blind human preference. On VoiceArena's Hindi leaderboard, filtered to real-time models and checked by Maya Research on October 27, 2026:
| Rank | Model | Elo | 95% CI |
|---|---|---|---|
| 1 | Cartesia Sonic-3.6 | 1094 | ±13 |
| 2 (2 to 4) | Maya Research Maya-2-Native | 1047 | ±11 |
| 3 (2 to 5) | Cartesia Sonic-3.5 | 1035 | ±12 |
| 4 (2 to 5) | Smallest AI Lightning 3.1 Pro | 1024 | ±12 |
| 5 (3 to 5) | Murf AI Falcon 2 | 1021 | ±14 |
| 7 (6 to 7) | Sarvam AI Bulbul-V3 | 957 | ±9 |
| 8 | xAI Grok TTS | 890 | ±9 |
| 9 | Microsoft Azure Dragon HD Omni | 855 | ±9 |
Bradley-Terry Elo from blind listener votes; parentheses show published rank intervals. Ranks two to five overlap, so the gap between Maya and the models just below it is not decisive. Some coverage has described Maya 2 as first for Hindi; the live check above shows second among real-time models. Scores shift as votes are added.
Intelligibility. Maya Research generated 20.3 hours of speech from 6,000 held-out sentences, 500 per language across 31 categories from conversation to addresses and tongue twisters, and had three speech recognizers transcribe it. The lower the character error rate (CER), the more of the speech the recognizer understood:
| Language | CER, held-out ASR | CER, Qwen3-ASR | Voice identity |
|---|---|---|---|
| Hindi | 1.57% | 2.05% | 99.6% |
| Marathi | 1.54% | 2.09% | 99.1% |
| English | 1.43% | 1.74% | 99.6% |
| Kannada | 2.23% | 2.72% | 97.9% |
| Gujarati | 2.26% | 3.76% | 98.9% |
| Telugu | 2.53% | 2.64% | 99.4% |
| Punjabi | 2.66% | 3.32% | 99.1% |
| Bengali | 2.68% | 2.83% | 98.5% |
| Odia | 3.54% | 4.33% | 97.4% |
| Malayalam | 4.03% | 4.20% | 99.4% |
| Tamil | 4.33% | 4.55% | 98.1% |
| Assamese | 5.54% | 6.59% | 96.4% |
Maya Research's release measurements, scorer v1. The held-out recognizer was not used in training or checkpoint selection. Voice identity is how often a speaker classifier picked the intended voice among the 20.
Across the full benchmark, outputs that two judges agreed had failed (looping, large insertions, very high error) came to 0.15%, and the intended voice was recognized 98.6% of the time. These are automated measurements: the card is explicit that it reports no human naturalness study and no controlled comparison with other models.
What it costs to run
| Serving measurement (one H100 NVL) | Result |
|---|---|
| Concurrent streams | 224 |
| Aggregate throughput | 323 seconds of audio per second |
| Time to first audio, steady state, p50 / p90 / p99 | 111 / 134 / 242 ms |
| Time to first audio, single stream | about 14 ms |
| GPU memory, 1 seat / 32 seats (steady) | 7.7 GB / 15.8 GB |
One seat needs about 9.2 GB at load, so a 12 GB GPU is a realistic floor for a single stream. The 224-stream run on one H100 is the more interesting number for anyone running a voice agent or IVR at scale: at a time-to-first-audio of around 110 ms, a single card can serve a large call center's worth of concurrent conversations.
Where it struggles
- Assamese. No built-in voice has an hour of Assamese training audio, so every Assamese output is cross-lingual and some drift toward a Bengali-like sound.
- Very short turns. One- or two-word replies are less stable in voice; a slightly longer phrase helps.
- URLs, symbols and code-mixing. Web addresses had a 29% character error rate on one judge. Listen to your own names, numbers and links before relying on them.
- Cloning quality is unmeasured. Similarity is approximate, and the release reports no cloning benchmark.
- Emotion tags such as laughs and whispers are best effort.
- No watermark. Consent, disclosure and abuse monitoring are the deployer's job.
Try it
On Linux with an NVIDIA GPU (CUDA 12.8) and Python 3.10 to 3.12:
pip install "maya3[server] @ git+https://github.com/MayaResearch/maya2-native" \
--extra-index-url https://download.pytorch.org/whl/cu128
from maya3 import Maya3TTS
tts = Maya3TTS.from_pretrained("maya-research/maya2-native", max_seats=1)
audio = tts.generate("नमस्ते, आपका ऑर्डर कल पहुँच जाएगा।", lang="hi",
voice="Ahana", normalize=True)
The package is still named maya3 internally. Samples in all twelve languages are on the model card, and the server, streaming and cloning examples are in the GitHub repository.
Our read
A five-person lab built on an open language model and codec and placed second in its home language against much larger companies. Data and tuning for a specific population are a real moat.
Banks, telecoms and government services often cannot send customer audio abroad. A strong Apache-2.0 model that runs on one GPU inside their own walls changes what they can deploy.
If the next users talk rather than type, the speech layer decides who can use AI at all. The test Reddy sets, whether his mother could use it, is the right benchmark.
The open questions are the ones the card itself raises: how it sounds to native listeners in each language, not just to recognizers; how well clones hold up; and how responsibly cloning is used without a watermark. For a broader view of the labs building models like this, see our Neolab List, and for why small, self-hosted models matter, What a 4GB Model Is For.
Before you deploy it
- Have native speakers listen to your real scripts in each language you will ship, especially names, numbers and links.
- Check which built-in voices were trained on your languages; others are cross-lingual and may carry an accent.
- Get explicit consent for any voice you clone, and tell callers they are hearing synthetic speech.
- Put authentication in front of the server; it has none built in.
- Size
max_seatsto your GPU memory and measure latency on your own network, not the benchmark's.
Sources: Maya Research, Maya 2 Native model card and inference code (release 1.0, October 25, 2026; benchmarks, architecture, limitations and team); VoiceArena Hindi TTS leaderboard, as reported on the model card for October 27, 2026; Distractify, August 19, 2026, and Analytics Insight (founder background); South Park Commons (founding year and location). Snack on AI, Inside Maya, India's Open Voice Model, the Small Model Is the Slow One, October 27, 2026. Benchmark figures are Maya Research's own unless stated.