README.md

metadata

title: Maya 2 Native TTS
emoji: 🗣️
colorFrom: blue
colorTo: pink
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
python_version: '3.10'
short_description: Streaming multilingual TTS with zero-shot voice cloning
startup_duration_timeout: 1h

Maya 2 Native — multilingual streaming TTS

Demo of maya-research/maya2-native: a 2.5B-parameter text-to-speech model covering 12 Indic languages + English (Hindi, English, Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu) with zero-shot voice cloning from a 3–10 s reference clip. Apache-2.0. Output is 24 kHz mono audio.

• Named voices: 20 trained voices ship with the release — pick one from the dropdown, or leave "(none)" for the plain (unclaimed) voice.
• Voice cloning: upload a 3–10 s clip plus its exact transcript; the same speaker reads your text — including in a different language (cross-lingual cloning).
• Text normalisation (on by default): expands digits, currency, dates, URLs and ALL-CAPS names into spoken words in the target language.

This Space vendors the maya3 inference engine (maya3/, tts/, serve/, scripts/, configs/) under its Apache-2.0 license. Model weights (Mimi codec, ECAPA speaker encoder) are downloaded from the Hugging Face cache on first use.

Examples

The example rows are the showcase texts from the model card, one per supported language, each paired with its native voice. The bundled samples/hi_ritvik.wav is a reference clip from the model repository you can use to try voice cloning (its transcript is in the model card).