Maya Research releases self-hosted streaming speech for 11 Indian languages and English

Maya Research released self-hosted streaming text-to-speech for 11 Indian languages and English.

Maya Research has released Maya 2 Native, a text-to-speech model that developers can run on their own infrastructure. Its October release pairs downloadable weights with inference code under Apache 2.0. The model supports Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu and English. The launch gives teams building phone services, narration and accessibility tools another way to generate speech for several Indian languages from one deployment. The model card dates the 1.0 checkpoint to October 25, with a separate release post describing how to use it. [Model card](https://huggingface.co/maya-research/maya2-native)

What the download provides

The checkpoint produces 24 kHz mono audio and ships with 20 preset voices. It can stream audio during synthesis or use three to ten seconds of reference speech to reproduce aspects of a speaker’s voice. The inference package exposes a Python interface, a speech endpoint compatible with the OpenAI API’s shape and a WebSocket for streaming. The project retains the package name `maya3` and the class `Maya3TTS` even though the public checkpoint is called Maya 2 Native. Teams reading examples or tracking versions should keep the two names separate. [Model card](https://huggingface.co/maya-research/maya2-native) · [Inference repository](https://github.com/MayaResearch/maya2-native)

Maya 2 Native occupies the speech generation stage of a voice pipeline. A caller’s audio must first reach a recognition component, and the application or a language model decides what text the system should say. The rest of the application manages when to begin playing that speech and when to stop if the caller interrupts. This is one reason a synthesizer’s streaming interface and a completed phone call measure different things. A model can begin returning audio early while the overall system still feels slow because recognition, response generation or network transport takes too long. [Model card](https://huggingface.co/maya-research/maya2-native)

Where the deployment tradeoffs show up

The published setup guidance starts with Linux and an NVIDIA GPU using a driver compatible with CUDA 12.8. Maya Research advises starting with one serving slot and reports a loading peak of about 9.2 GB plus CUDA overhead, suggesting more headroom on a card with at least 12 GB of memory. These are setup figures for a documented configuration. For an IVR team, the meaningful questions are how quickly a caller hears the first usable audio, whether a line stays responsive under concurrent calls and whether the speech remains intelligible when the application reads a real account balance or a date. [Maya Research announcement](https://huggingface.co/blog/maya-research/maya2-native)

The source also makes a useful distinction between language support and language quality. The model card reports automated checks that use speech recognition and speaker embeddings, with limits on how far those results generalize across dialects or speakers. A rollout for a regional service could start with prompts native speakers actually hear, including local place names, code-mixed phrases, prices and appointment times. Testing those cases also helps isolate whether an error comes from text normalization, pronunciation or the voice selected for the prompt. [Model card](https://huggingface.co/maya-research/maya2-native)

The downloadable weights give operators more choice over where inference runs. Voice cloning adds a separate governance decision because reference audio may identify a real speaker. Maya Research’s model card calls for permission to use reference voices and disclosure of synthetic speech, especially in calls and media. For products that need only a consistent synthetic voice, the presets give teams a path to evaluate the model without collecting someone’s recordings. [Model card](https://huggingface.co/maya-research/maya2-native)