# Maya 2 Native

> Open-weight streaming text-to-speech for 11 Indian languages and English, with 20 built-in voices, zero-shot voice cloning and 24 kHz mono output. Weights and inference code: Apache-2.0.

Canonical model: maya-research/maya2-native. Public aliases: Maya2 Native, Maya-2-Native. The working Python package remains maya3 and the loader is maya3.Maya3TTS.from_pretrained. The model is TTS, not speech recognition or a conversational LLM.

Languages: as (Assamese), bn (Bengali), en (English), gu (Gujarati), hi (Hindi), kn (Kannada), ml (Malayalam), mr (Marathi), or (Odia), pa (Punjabi), ta (Tamil), te (Telugu). Reference engine: Linux, Python 3.10–3.12, NVIDIA CUDA 12.8. Start with max_seats=1. Built-in checkpoint voices differ from hosted Maya API voices. Documentation reviewed November 3, 2026.

## Model and integration

- [Model card, raw Markdown](https://huggingface.co/maya-research/maya2-native/resolve/main/README.md): Full model description, examples, voice coverage, reported benchmarks, limitations, credits and component licences.
- [Agent integration guide](https://huggingface.co/maya-research/maya2-native/resolve/main/AGENT_GUIDE.md): Exact loader, installation, language codes, streaming formats, endpoints and interface distinctions.
- [Capability manifest, JSON](https://huggingface.co/maya-research/maya2-native/resolve/main/model-discovery.json): Machine-readable model identity, languages, voices, audio contract, runtime requirements and source links.
- [Inference-engine README](https://raw.githubusercontent.com/MayaResearch/maya2-native/main/README.md): Python, HTTP and WebSocket setup.
- [Engine documentation](https://raw.githubusercontent.com/MayaResearch/maya2-native/main/docs/ENGINE.md): Memory, batching, concurrency and engine configuration.
- [Local server protocol](https://raw.githubusercontent.com/MayaResearch/maya2-native/main/maya3/server.py): HTTP fields, audio formats, WebSocket messages, readiness and errors.
- [Built-in voice metadata](https://huggingface.co/maya-research/maya2-native/resolve/main/voices/voices.json): Exact voice names and per-voice training-language coverage.
- [Text normalization](https://raw.githubusercontent.com/MayaResearch/maya2-native/main/maya3/text/README.md): Spoken forms for numbers, dates, currency, URLs and other text.

## Optional

- [Human-readable model page](https://huggingface.co/maya-research/maya2-native): Listen to twelve speech samples and watch the model video.
- [Hosted Maya Pipecat integration](https://raw.githubusercontent.com/MayaResearch/pipecat-maya/main/README.md): Hosted API integration; use its own voice IDs and protocol, not the local checkpoint's contract.
- [Maya Transcribe](https://huggingface.co/maya-research/maya-transcribe): Separate Hindi/Hinglish speech-to-text release.