---
title: Praxy Voice
emoji: 🎤
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 5.38.2
app_file: app.py
pinned: false
license: apache-2.0
short_description: Hi · Te · Ta · En TTS · code-mix · voice cloning
models:
  - Praxel/praxy-voice-r6
  - ResembleAI/chatterbox
  - ai4bharat/IndicF5
---

## Required Space secrets
- `MODAL_TOKEN_ID` — get from Modal dashboard
- `MODAL_TOKEN_SECRET`
- `ANTHROPIC_API_KEY` — for the codemix transliteration preprocessor

## Hardware
Free CPU is sufficient — this Space dispatches all GPU work to Modal.

## Custom domain (voice.praxel.in)
HF Space settings → Custom Domain → add `voice.praxel.in`, then add CNAME at praxel.in DNS pointing to the Space URL.

# Praxy Voice — Gradio Demo

Live demo of **Praxy Voice**, an open-source commercial-class Indic TTS system. Type Telugu, Tamil, or Hindi text and hear it spoken in a natural voice — with your own uploaded reference clip, or a bundled default.

## How it works

Three branches behind a single interface:

| Language | Path |
|---|---|
| Telugu | R6 LoRA on Chatterbox + BUPS romanisation + Hi-proxy `language_id` |
| Tamil  | R6 LoRA on Chatterbox + BUPS romanisation + Hi-proxy `language_id` |
| Hindi  | **Vanilla** Chatterbox (LoRA not needed — Chatterbox natively covers Hi) |

All three share the **voice-prompt recovery** recipe: an 8–11 s same-language reference audio clip + Config B sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1).

See the companion paper for the full methodology: *"Praxy Voice: Voice-Prompt Recovery + BUPS for Commercial-Class Indic TTS"* ([arXiv:2604.25441](https://arxiv.org/abs/2604.25441)).

## Local run

```bash
cd gradio_demo
pip install -r requirements.txt
python app.py
```

## Files

- `app.py` — Gradio interface + language router + inference
- `requirements.txt` — Python deps
- `data/` — bundled default reference clips (place your own here before deploy)

## Defaults and limits

- Default reference clips: Sarvam Bulbul Te/Ta + Cartesia Sonic-3 Hi (8–11 s each).
- Max input text length: not enforced in UI but practically ~400 chars.
- Hardware: runs on ZeroGPU (T4-class; ~2–5 s per generation).

## License

Apache-2.0. See `LICENSE` in the model repo `Praxel/praxy-voice-r6`.