Voice Clone Lab

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.