# Voice Clone Lab

> Locally fine-tune Qwen3-TTS on your own voice from a few minutes of audio and generate speech via CLI or web UI — fully on-device with no data leaving your machine.

Voice Clone Lab is an open-source Python toolkit by tetsuo-ai that lets you clone a voice from 5–15 minutes of clean audio and generate speech from text entirely on your own GPU. It wraps Alibaba's Qwen3-TTS model in a guided fine-tuning pipeline accessible through both a command-line interface and a Gradio web UI, with all audio, transcripts, and trained models staying local.

## What It Is

Voice Clone Lab is a local voice-cloning and text-to-speech fine-tuning pipeline built on top of Qwen3-TTS (Apache-2.0). It accepts audio from a file upload, microphone recording, or a YouTube/URL link, then walks the user through extraction, cleaning, splitting, transcription, proofreading, and training — producing a personal voice model organized by speaker name. A zero-shot mode lets users try a voice instantly from a single clip without any training.

## How the Pipeline Works

The tool structures the workflow into discrete, repeatable steps that can be run individually or chained together:

- **Data ingestion**: `vcl run --speaker <name> --input <file or URL>` handles extraction, cleaning, splitting, and transcription in one command.
- **Transcript review**: An editable TSV file lets users fix transcription errors before training — the README notes bad transcripts are the primary cause of pronunciation problems.
- **Training**: `vcl prepare` adds Qwen audio codes; `vcl train` fine-tunes the model. The web UI exposes hyperparameter controls and a live training log.
- **Generation**: `vcl generate` produces WAV output from the newest checkpoint; a `--zeroshot` flag enables instant cloning from a reference clip; a `--spec` flag supports batch generation from a JSON list.
- **Organization**: All data lives under `data/voices/<name>/` and `outputs/checkpoints/<name>/`, with epoch snapshots kept independently so users can compare and select the best-sounding checkpoint.

## Hardware Requirements and Tradeoffs

The project targets Linux machines with NVIDIA CUDA GPUs. Training requires a minimum of 24 GB VRAM (RTX 3090/4090 class at batch size 2), while generation-only use works on cards with around 8 GB VRAM. Python 3.10 or later is required, along with `ffmpeg` and `git`. Optional extras — WhisperX for word-level timestamps, yt-dlp for URL sources, WebRTC VAD for better audio chunking, and mild denoising — are installable as pip extras. FlashAttention is optional; the tool falls back to `sdpa` automatically.

## Architecture and Dependencies

Voice Clone Lab is built on Qwen3-TTS (Apache-2.0) by Alibaba's Qwen team, with faster-whisper for transcription, Gradio for the web UI, and yt-dlp for URL-based audio sourcing. The repository applies a vendor patch to Qwen3-TTS via a `patches/` directory managed by `vcl setup`. The base model weights (~4 GB) are downloaded during setup. Source layout follows a standard Python package structure under `src/voice_clone_lab/`, with a single `config/default.yaml` as the configuration source.

## Current Status

The repository was created on July 18, 2026, and last updated on July 19, 2026, indicating it is a very recently published project. It is licensed under Apache-2.0 and hosted publicly on GitHub under the tetsuo-ai organization. The project includes a CPU-only test suite and an `AGENTS.md` conventions document, suggesting active development infrastructure is in place from the outset.

## Features
- Clone a voice from 5–15 minutes of audio
- Fine-tune Qwen3-TTS locally on your own GPU
- CLI and Gradio web UI interfaces
- Audio ingestion from file upload, microphone, or YouTube/URL link
- Guided pipeline: extract, clean, split, transcribe, proofread, train
- Editable transcript review via TSV before training
- Zero-shot voice cloning from a single reference clip
- Batch speech generation from JSON spec file
- Speaker-organized data and checkpoint management
- Epoch snapshot comparison for best checkpoint selection
- Fully local — no data leaves your machine
- Optional WhisperX, yt-dlp, WebRTC VAD, and denoising extras

## Integrations
Qwen3-TTS, faster-whisper, Gradio, yt-dlp, WhisperX, WebRTC VAD, PyTorch, CUDA, ffmpeg

## Platforms
LINUX, CLI, API

## Pricing
Open Source

## Version
main

## Links
- Website: https://github.com/tetsuo-ai/voice_clone_lab
- Documentation: https://github.com/tetsuo-ai/voice_clone_lab#readme
- Repository: https://github.com/tetsuo-ai/voice_clone_lab
- EveryDev.ai: https://www.everydev.ai/tools/voice-clone-lab