Practical overview
When it fits
Use it when you can run a local or server-side Python training workflow and want one project for audio preparation, Qwen3-TTS fine-tuning and checkpoint inference.
Who it is for
- Developers and practitioners preparing speech datasets for Qwen3-TTS fine-tuning.
- Teams or individuals training custom voices and testing speech synthesis from checkpoints.
Documented uses
- Split raw audio, transcribe and clean it with ASR, and tokenize the resulting dataset.
- Fine-tune Qwen3-TTS models using the project’s training workflow.
- Run inference with a trained checkpoint to try a custom voice.
Requirements & limits
The repository is public and licensed under Apache-2.0. The README documents Docker and Python virtual-environment setup.
- The recommended training setup calls for an NVIDIA GPU with at least 16 GB VRAM, 32 GB RAM and 50 GB of free SSD space.
- The README recommends Linux and CUDA 12.4 or later, with CUDA 12.8 or later suggested.
- Python is specified as 3.10 or later; the project identifies PyTorch as part of its stack.
- Direct Windows Python setup is not actively maintained; Windows GPU guidance covers WSL2 and Docker Desktop, with a stability caveat for Docker Desktop.
- The repository provides presets for Qwen3-TTS 0.6B and 1.7B model variants.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare, train and test a voice with the CLI
Use the documented CLI stages to prepare speaker audio, train a Qwen3-TTS model and run inference on a checkpoint.
Before you start
- A Linux training environment with Python 3.10 or later and CUDA 12.4 or later; CUDA 12.8 or later is recommended.
- An NVIDIA GPU with at least 16 GB VRAM; 24 GB is recommended for the 1.7B model, along with 32 GB RAM and 50 GB free SSD space.
- Raw WAV recordings placed in a speaker-named directory.
Steps
Place WAV files in a directory such as raw-dataset/my_speaker/.
Run the documented prepare command with the input directory and speaker name.
Run the documented train command with an experiment name, speaker name and epoch count.
Run the documented infer command with the resulting checkpoint, speaker and text.
Expected outcome
A trained checkpoint can be used for inference with the selected speaker and text.
Guide limitations
- The README’s hardware figures are recommendations for stable training, not a compatibility guarantee.
- For CustomVoice fine-tuning, the README specifies generating a speaker embedding after ASR and before training.
Set up the Gradio WebUI
Install the project and launch its browser interface for audio preparation, training and inference.
Before you start
- A Linux environment with Python 3.10 or later and CUDA 12.4 or later for the recommended training setup.
- An NVIDIA GPU with at least 16 GB VRAM, 32 GB RAM and 50 GB free SSD space are recommended.
Steps
Create and activate a Python virtual environment.
Install the project requirements and the documented Flash Attention version matching the CUDA and PyTorch setup.
Launch the interface with python src/webui.py.
Use the interface’s documented stages to upload and prepare audio, configure training, and load a checkpoint for inference.
Expected outcome
The Gradio interface provides browser-based access to data preparation, training monitoring and interactive inference.
Guide limitations
- Direct Windows Python setup is not actively maintained; the README recommends Docker for Windows users.
- The README does not claim every Windows Docker GPU setup will be stable.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- mozi1924/Qwen3-TTS-EasyFinetuning · Official homepage title · retained in its source language