qw3n-face

A local web UI for the Qwen3-TTS model family, built with NiceGUI.

Features

Tab	Description
Custom Voice	Generate speech using one of 9 built-in speaker personas with optional style instructions
Voice Design	Describe a voice in plain text and synthesise speech with it
Voice Clone	Upload a short reference clip and clone that voice onto new text
Batch	Queue multiple Custom Voice items and generate them sequentially with per-item progress
Personas	Save and manage named voice presets (speaker + language + instruction) for quick reuse

Models are loaded on demand and can be unloaded individually to free memory. Custom Voice and Voice Clone support both 0.6B and 1.7B checkpoints; Voice Design currently uses the 1.7B checkpoint only.

Additional runtime behavior:

• Choose the model size before loading when multiple checkpoints are available
• Choose the backend device before loading (cuda:0, mps, or cpu, depending on your machine)
• Loaded tabs show the active runtime as device / dtype
• On Apple Silicon, the app retries once in safer MPS float32 mode if generation fails with a probability-tensor stability error

Requirements

• Python 3.11+
• uv
• A machine with MPS (Apple Silicon), CUDA, or enough RAM for CPU inference

Installation

git clone https://github.com/AlapinEnjoyer/qw3n-face.git
cd qw3n-face
uv sync

Running

uv run python main.py

Then app should auto open itself on http://localhost:8080.

Models

Models are downloaded automatically from Hugging Face once requested in the app:

Key	Checkpoint
Custom Voice	Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice or Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice
Voice Design	Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
Voice Clone	Qwen/Qwen3-TTS-12Hz-1.7B-Base or Qwen/Qwen3-TTS-12Hz-0.6B-Base

Approximate checkpoint sizes vary by model variant; 0.6B models are substantially smaller than 1.7B models. Downloads are cached locally by Hugging Face after the first load.

Runtime Notes

• CUDA uses bfloat16
• MPS prefers float16, but the app can retry a failing model in float32 for stability
• CPU prefers bfloat16 and falls back to float32 if needed during model load
• If Apple Silicon generation still fails on MPS, switch the backend device to cpu before loading the model

TODOs

• Add automatic transcription of uploaded audio
• Add audio visualisation (waveform, spectrogram?)

Roadmap

• Add support for fine tuning