Maya Research releases Maya 2 Native, an Apache 2.0 streaming TTS model for 11 Indian languages and English
Maya Research has released Maya 2 Native, an open-weight text-to-speech model covering 11 Indian languages and English under Apache 2.0. It streams 24 kHz mono audio in 80-millisecond frames and supports cross-lingual voice cloning from a 3 to 10 second reference clip.
Maya Research has released Maya 2 Native, an open-weight text-to-speech (TTS) model that generates speech in 12 languages: 11 Indian languages plus English. The model is licensed under Apache 2.0 and supports real-time audio streaming and voice cloning, according to Maya Research's documentation, as reported by Analytics India Magazine.
Key specifications
- Languages: Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu.
- Audio output: 24 kHz mono, streamed in 80-millisecond frames, so applications can begin playback while generation continues.
- Voices: 20 built-in synthetic voices.
- Voice cloning: Uses a reference clip of around three to 10 seconds. A sample recorded in one supported language can be used to generate speech in another.
- License: Apache 2.0, allowing use and modification of the released model and code, subject to the license terms.
- Hardware: Local deployment requires an NVIDIA GPU. The documentation recommends at least 12 GB of GPU memory for a single-session setup.
Maya Research's documentation describes the release as "Streaming text-to-speech for 11 Indic languages and English."
Integration options
The release includes inference code and interfaces for Python, an OpenAI-compatible speech endpoint and WebSocket streaming. Developers can integrate the model into applications or deploy it on their own infrastructure. The target use cases named are voice assistants and regional-language AI services.
Evaluation
Maya Research evaluated the model on 6,000 text samples across the 12 languages, producing 20.3 hours of speech. Three automated speech-recognition systems assessed intelligibility, and character error rates varied by language and by evaluation system. The documentation reports lower error rates for several languages, including Hindi, Marathi and Gujarati. Assamese proved more challenging.
The model card does not report a human naturalness study or a controlled comparison with other TTS models. It also lists limitations with code-mixed text, URLs, numbers and some cross-lingual voice combinations. On cloning, the documentation notes that cloned voices approximate the reference, with similarity and accent varying across outputs.
Safety and consent
Maya Research recommends obtaining permission before cloning a voice and telling listeners when speech is synthetic, particularly in calls and media. The release does not include an audio watermark. Consent, disclosure and abuse monitoring are left to the developers who deploy the model.
What this means
Maya 2 Native gives developers a permissively licensed TTS option that covers 11 Indian languages in one model, with streaming output and an OpenAI-compatible endpoint that can simplify integration into existing voice stacks. The 80-millisecond frame size is relevant for conversational applications where playback must begin before synthesis finishes.
Quality relative to other TTS systems is not established by the published materials. Evaluation relied on automated speech recognition, with no human naturalness study and no head-to-head comparison, and the reported results vary by language, with Assamese weaker. Teams building for production should test the languages and text types they need, especially code-mixed text and numbers, which the model card flags as limitations.
Because there is no built-in watermark, anyone using the voice cloning feature takes on responsibility for consent and disclosure controls.
Related Articles
vLLM Semantic Router releases d3-ultra, a 27B open-weight multimodal decision model
vLLM Semantic Router has released d3-ultra, a 27B open-weight multimodal decision model. It reads text, images and video once and returns probabilities for every answer. The team reports a score of 87.2 on its SR-Text test.
Tencent Hunyuan open-sources Hy-SimulMT, a 30B-A3B speech-to-speech interpreter with about 3 seconds of lag
Tencent Hunyuan has open-sourced Hy-SimulMT, a 30B-A3B mixture-of-experts model that translates Chinese and English speech in real time while cloning the speaker's voice. Tencent reports about 3 seconds of StreamLAAL lag on the long-form RealSI benchmark, the lowest among the systems it compared. The weights are listed under Apache 2.0.
KRAFTON releases Raon-1 mini preview, a 28.8B-parameter MoE model with 4.3B active and 262,144-token context
KRAFTON AI has released Raon-1 mini preview, a Mixture-of-Experts language model trained from scratch with 28.8B total and 4.3B active parameters. It is Apache 2.0 licensed, supports 262,144 tokens of context, and ships with a speculative decoding draft model that KRAFTON says reaches up to 2.18× higher output throughput on a single NVIDIA B300.
Kimi K3.1 Pro launches with 1M-token context, 65.7% on Terminal-Bench 4.0; open weights due by Nov 4
Kimi K3.1 Pro, the first model in the K3.1 series, is a 3T-class open-weight model with a 1-million-token context window. It scores 65.7% on Terminal-Bench 4.0 and is available today on Kimi.com, Kimi Work, Kimi Code and the Kimi API. Full weights will be released by Nov 4.
Comments
Loading...