README.md

---
title: FireRedTTS3
emoji: 🔥
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 6.24.0
app_file: app.py
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
license: apache-2.0
short_description: Voice cloning, voice design and speech editing
models:
  - FireRedTeam/FireRedTTS3
---

# 🔥 FireRedTTS3

Interactive demo of [FireRedTeam/FireRedTTS3](https://huggingface.co/FireRedTeam/FireRedTTS3),
a unified speech generation and editing model (Qwen3-1.7B backbone + flow-matching
head over the continuous RedAE audio autoencoder, 24 kHz).

Three tabs:

| Tab | Checkpoint | What it does |
| --- | --- | --- |
| 🎙️ Voice Cloning | `fireredtts3_base` | Zero-shot cloning from a short reference clip — 24 languages and 21 Chinese dialects |
| 🎨 Voice Design | `fireredtts3_instruct` | Creates a voice from a natural-language description, no reference audio |
| ✂️ Speech Editing | `fireredtts3_instruct` | Semantic edits (insert / delete / substitute words) and acoustic edits (speed, pitch, volume) |

Text normalization runs locally via `wetext` (Chinese / English); automatic language
routing uses fastText `lid.176`. The optional LLM-based normalizer from the upstream
repo is disabled here (it requires external API credentials).

## Credits

- Model and inference code: [FireRedTeam/FireRedTTS3](https://github.com/FireRedTeam/FireRedTTS3), Apache-2.0.
- `examples/en_prompt.wav`: from [OpenBMB/VoxCPM](https://github.com/OpenBMB/VoxCPM) (`examples/example.wav`), Apache-2.0.
- `examples/zh_prompt.wav`: from [FireRedTeam/FireRedTTS2](https://github.com/FireRedTeam/FireRedTTS2) (`examples/chat_prompt/zh/S2.flac`), Apache-2.0.

Voice cloning is provided for academic research purposes only — do not use it for
impersonation or any illegal activity.