Editorial illustration for Maya Research Releases Open-Weight Speech Model With 12 Languages and 20 Voices
Maya 2 Native: Open-Weight Speech Model, 12 Languages
Maya Research put out a new text-to-speech model this week called Maya 2 Native, and it covers 11 Indian languages plus English under an Apache 2.0 license. That license matters: developers can take the model, modify it, and deploy it on their own infrastructure without the usual restrictions attached to proprietary voice AI.
The model handles Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, and English, generating 24 kHz mono audio streamed in 80-millisecond frames. That frame size lets an app start playing audio before generation finishes, which is the difference between a voice assistant that feels instant and one that makes you wait. Maya Research built in 20 synthetic voices developers can pick from out of the box, and the model also does voice cloning from a short reference clip, roughly three to 10 seconds long.
The cloning feature has a cross-language trick built in: feed it a sample in one language, and it can generate speech in another supported language while keeping the voice characteristics. Maya Research shipped inference code alongside it, along with a Python interface, an OpenAI-compatible speech endpoint, and WebSocket streaming support.
Maya Research has released Maya 2 Native, an open-weight text-to-speech model that generates speech in 12 languages, including 11 Indian languages and English. The model supports real-time audio streaming and voice cloning, allowing developers to build multilingual voice applications.
Why this matters
The Apache 2.0 license is the real story here, not the voice count. Most commercial TTS APIs charge per character and lock you into their voice library. Maya 2 Native hands developers the weights, which means self-hosting, fine-tuning, and no per-call billing eating into margins on a voice app. For founders building in India or other markets where 11 of these 12 languages matter, that changes the unit economics of launching a product versus renting one from ElevenLabs or Google.
The cross-lingual voice cloning feature deserves scrutiny, though. A three-to-ten-second clip generating speech in a language the speaker never spoke is a strong claim, and accent fidelity, prosody, and cultural nuance in code-switched regions tend to break these systems in practice. We'd want to hear it on Tamil or Bengali samples before taking the claim at face value.
Still, for researchers studying low-resource language TTS, having an open model trained on this language set to study and modify is worth more than another closed API demo. Watch for independent benchmarks on the Indian language outputs specifically.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Newest AI Quest Maps the Human Brain, Builds on Fruit Fly Brain Map
From Archives iLands AI agents pitch WIRED for fact-checking, editor Vojdani says they spam
Common Questions Answered
What languages does Maya 2 Native support and under what license is it released?
Maya 2 Native supports 12 languages total: 11 Indian languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu) plus English. The model is released under an Apache 2.0 license, which allows developers to freely modify and deploy it on their own infrastructure without the restrictions typically associated with proprietary voice AI services.
How does Maya 2 Native's open-weight model differ from commercial TTS APIs like ElevenLabs or Google?
Maya 2 Native provides developers with the actual model weights, enabling self-hosting, fine-tuning, and elimination of per-call billing charges that commercial APIs typically impose. This fundamentally changes the unit economics for voice applications, as developers are no longer locked into proprietary voice libraries or charged per character like traditional commercial TTS services.
What real-time capabilities does Maya 2 Native offer for developers building voice applications?
Maya 2 Native supports real-time audio streaming at 24 kHz mono quality and includes voice cloning functionality, allowing developers to build multilingual voice applications with custom voice options. These capabilities enable developers to create responsive voice experiences without relying on external APIs or services.
Why is the Apache 2.0 license significant for founders building products in Indian markets?
The Apache 2.0 license is significant because it eliminates licensing restrictions and per-call billing costs, making it economically viable for founders to launch voice products rather than renting capabilities from expensive commercial providers. For markets where 11 of the 12 supported languages are directly relevant, this open-weight approach dramatically improves product margins and reduces dependency on third-party services.
Further Reading
- Maya TTS: streaming Indic + English text-to-speech - GitHub
- Maya 2 Native multilingual streaming TTS model - Hugging Face
- Maya 2 TTS: Eleven Indian Languages, One Awkward Folder Name - DEV Community
- Maya 2 Native TTS on Apple Silicon with MLX - GitHub
