---
title: M5-TTS
emoji: 🔊
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 6.15.1
app_file: app.py
python_version: "3.10"
startup_duration_timeout: 30m
suggested_hardware: zero-a10g
short_description: Reference-conditioned multilingual M5 speech synthesis
---

# M5-TTS

Reference-conditioned multilingual text-to-speech synthesis with a 16-codebook,
12.5 Hz Qwen speech codec. Target and reference transcripts use the verified
`m5_tn/m5_infer.py` frontend with WeTextProcessing enabled by default.
The codec is loaded from the pinned official
`Qwen/Qwen3-TTS-12Hz-1.7B-Base` revision at startup.

## Usage

1. Upload or record a clean reference audio clip.
2. Enter the exact transcript of the reference.
3. Enter up to 250 characters of target text.
4. Select language and optional Cantonese (`yue`) or Sichuan (`chuan`) accent.
5. Generate and download the resulting audio.

The app requires GPU hardware. It is packaged for Hugging Face ZeroGPU and uses
one GPU allocation for the complete request.