# Voice Clone Lab Voice Clone Lab clones a voice from a few minutes of audio and generates speech locally on your own GPU. Give it 5 to 15 minutes of clean speech and it fine-tunes [[Qwen3-TTS]] into a personal voice model you drive from a CLI or a web UI. By tetsuo-ai, Apache-2.0, [[Python]]. ## Consent, first The README leads with this and so should the note: > Use this project only for **your own voice** or a voice where you have **explicit written permission** from the speaker. Do not use it for impersonation, fraud, bypassing consent, harassment, or cloning public figures or private people without permission. The same rule applies to URLs: only download content you have the rights to use. Voice cloning is dual-use in a way that most tools in this space aren't, and the fact that this one puts the constraint above the install instructions rather than in a footer is worth noting. ## The pipeline One command runs the whole thing: ```bash vcl run --speaker alex --input ~/recordings/alex.wav # or a YouTube/URL source, if you have the rights to it ``` That extracts, cleans, splits, transcribes and builds the dataset. Then: ```bash vcl transcribe --speaker alex --review # editable TSV vcl transcribe --speaker alex --apply-review --force vcl prepare --speaker alex # add Qwen audio codes vcl train --speaker alex # fine-tune vcl generate --speaker alex --text "Hello, this is my cloned voice." ``` **The proofreading step is the one to not skip.** The README says bad transcripts are the number one cause of weird pronunciation, and it builds an editable review table specifically so you fix them before training. Two minutes of correction against a fine-tune you'd otherwise have to redo. Everything is keyed off the speaker name, so there's no per-voice config to edit. Audio, chunks, transcripts and reference clips land under `data/voices/<name>/`, checkpoints under `outputs/checkpoints/<name>/`. There's also a **zero-shot mode** for trying a voice from a single clip with no training at all, which is the right way to decide whether a source recording is worth the GPU time. ## What it needs - Linux with an NVIDIA CUDA GPU - **Training**: 24 GB VRAM minimum (3090/4090 class, `--batch-size 2`), 32 GB+ comfortable - **Generation only**, using someone else's checkpoint: any ~8 GB card - Python 3.10+, `ffmpeg`, `git` Optional extras pull in WhisperX for word timestamps, `yt-dlp` for URL sources, WebRTC VAD for better chunking, and mild denoising. A [[Gradio]] web UI (`vcl ui`) mirrors the CLI across four tabs: setup, data, train, generate. ## Versus Voice Clone Studio I have a note on [[Voice Clone Studio]], and these are close enough to confuse. Voice Clone Studio is the **broad** one: multiple engines (Qwen3-TTS, VibeVoice, LuxTTS), voice design from text descriptions, multi-speaker conversation up to 90 minutes, emotion presets, preset premium voices. A studio, as the name says. Voice Clone Lab is the **narrow** one: a single engine and a single job, cloning one voice properly, with the dataset pipeline treated as the main event rather than a feature. The transcript review table and the "don't skip this" warning are the tell. Pick by what you're doing. Exploring what's possible across engines, take the studio. Producing one high-quality personal voice model, the lab's opinionated pipeline is more likely to get you there. ## Why I keep this note **It's adjacent to [[Knowii Voice AI]] from the opposite direction.** KVA is speech in: transcription, local, private, my own hardware. This is speech out, with the same commitments. Same stack neighbourhood, same local-first argument, same class of user who doesn't want their voice on someone's server. **Fully local is the entire value proposition.** Commercial voice cloning means uploading a recording of yourself to a company and trusting their retention policy. Your voice is biometric data you can't rotate after a breach. A pipeline where audio, transcripts and models never leave the machine is not a nice-to-have here, it's the reason to use it. **The dataset step is where quality actually lives.** Not the model, not the hyperparameters. Clean audio, correct transcripts, sensible chunking. The tool is built around that belief and says so out loud, which matches what I keep finding everywhere else: the boring preparation step is where results are won, and it's the step everyone wants to skip. Worth trying on my own voice before it's ever worth an opinion about the output quality. ## References - [tetsuo-ai/voice_clone_lab](https://github.com/tetsuo-ai/voice_clone_lab) — source - [Qwen3-TTS](https://github.com/QwenLM/Qwen3-TTS) — the model it fine-tunes ## Related - [[Qwen3-TTS]] - [[Voice Clone Studio]] - [[Voice Cloning]] - [[Knowii Voice AI]] - [[Whisper]] - [[Gradio]] - [[Python]] - [[FFmpeg]] - [[Qwen]] - [[Open Source]]