Maya Research has opened its voice model to developers. Maya 2 Native is an open-weight text-to-speech system that generates speech in 12 languages, including 11 Indian languages and English. It also supports real-time audio streaming and cross-lingual voice cloning — two features that turn a language model into something people can actually hear.
The supported languages are Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu. Maya Research describes the release as “Streaming text-to-speech for 11 Indic languages and English,” a direct pitch to developers building multilingual voice applications rather than another demo that stops at English.
Maya 2 Native generates 24 kHz mono audio and streams it in 80-millisecond frames for continuous playback. That setup supports applications that need speech as text arrives, instead of waiting for an entire response to finish rendering before anyone hears it.
What Developers Get
The release includes inference code and Python interfaces, an OpenAI-compatible speech endpoint, and WebSocket streaming. Developers can integrate the model into applications or deploy it on suitable infrastructure, while the Apache 2.0 licence allows them to use and modify the released model and code subject to the licence terms.
They also get 20 built-in synthetic voices for different applications. A developer can generate speech, imitate a reference voice, or use a sample recorded in one language to produce speech in another supported language.
Voice cloning works from a reference audio clip of around three to 10 seconds. The documentation warns that cloned voices only approximate the reference, with similarity and accent varying across outputs — a useful disclaimer in a field that often treats “clone” as a synonym for “perfect copy.”
Running Maya 2 Native locally requires an NVIDIA GPU, and the documentation recommends at least 12 GB of GPU memory for a single-session setup. That makes the model accessible to developers with suitable hardware, but not magically free of hardware requirements.
Language Coverage Comes With Caveats
Maya Research evaluated the model on 6,000 text samples across 12 languages and generated 20.3 hours of speech. Three automated speech-recognition systems assessed intelligibility, with character error rates varying by language and evaluation system.
The documentation reports lower error rates for several languages, including Hindi, Marathi, and Gujarati, while Assamese presented greater challenges. The model card also identifies problems with code-mixed text, URLs, numbers, and some cross-lingual voice combinations.
The evaluation does not include a human naturalness study or a controlled comparison with other text-to-speech models. That leaves an important question unanswered: how natural does Maya 2 Native sound to people, rather than to automated speech-recognition systems?
The release also arrives as researchers examine why AI systems treat languages differently before speech generation even begins. A study published by the Coalition for Language Equity in AI Research found that models can require far more tokens for some languages than others, creating what researchers call a hidden “tokenizer tax.”
Among the 20 most popular languages on the internet, the most expensive used between 1.5 and 5 times more tokens than the cheapest. When researchers expanded the analysis to 200 languages, the token cost gap in one model grew by 17 times, raising latency and reducing effective context capacity where compute and connectivity are already limited.
The gap is not only about how much digital content a language has. The study found that script plays a major role: Asturian benefits from the Latin-script advantage and representation from related Romance languages such as Italian, while Burmese, Tibetan, Tifinagh, Khmer, Lao, and Ol Chiki faced the greatest disadvantage across nearly all models in the analysis.
Dheemanth Reddy, CEO of Maya Research, said, “Token prices are falling very quickly, so the direct cost impact should become less severe over time, and better multilingual and synthetic data should improve model quality. But actually fixing the tokenization gap will still require labs to deliberately design future tokenizers to handle these languages and scripts better.”
That puts Maya 2 Native in a useful but demanding position. Its language coverage addresses a real access problem, while its limitations show that supporting a language means more than adding it to a list.
The company recommends obtaining permission before cloning a voice and disclosing to listeners that generated speech is synthetic, especially in calls and media. Maya 2 Native does not include an audio watermark, so developers must implement their own consent, disclosure, and abuse-monitoring measures.
Open weights and an Apache 2.0 licence give developers room to build, modify, and deploy. The harder work is making sure those systems remain intelligible, transparent, and safe across all 12 languages — not just the ones that already receive the best treatment from AI infrastructure.
Based on
- Maya Research Releases Open-Weight Speech Model Supporting 12 Languages — analyticsindiamag.com
- The Tokenizer Tax: Why non-english AI users pay more for tokens – Rest of World — restofworld.org
- Al Jazeera – Breaking News, World News and Video from Al Jazeera — aljazeera.com
- Teasa Adds Multilingual Character Voices to Its Interactive Storytelling Platform – USA Today — usatoday.com
- AI voice cloning helps woman with ALS get her voice back — nbcnews.com
