Voice Clone Lab: Free Local Voice Cloning With Qwen3-TTS

Voice Clone Lab is a free, local-first, self-hosted voice cloning pipeline that fine-tunes Qwen3-TTS from a few minutes of clean speech and generates new speech from text.

It runs entirely on your own Linux with an NVIDIA CUDA GPU. No audio or transcripts leave your device.

What Makes Voice Clone Lab Different

Many local voice-cloning tools generate speech directly from a short reference clip. Voice Clone Lab adds a complete fine-tuning workflow around Qwen3-TTS.

Zero-shot mode accepts one reference recording and generates speech immediately. The training workflow processes several minutes of speech, creates a speaker-specific dataset, and saves reusable checkpoints under that speaker’s name.

The project also treats transcript correction as a normal stage of training. Automatic transcripts appear in an editable TSV file or web table before dataset preparation. Corrections feed back into the dataset before the model creates audio codes and starts fine-tuning.

This structure works best for repeated narration and voice production. A saved speaker checkpoint becomes a reusable local model rather than a reference clip that must accompany every generation request.

Key Features

• Fine-tunes Qwen3-TTS into a reusable personal voice model from 5–15 minutes of clean speech.
• Processes recordings through audio extraction, cleaning, splitting, transcription, transcript review, and dataset preparation.
• Accepts microphone recordings, local audio and video files, and supported media URLs.
• Displays automatic transcripts in an editable TSV file or web table before training.
• Stores audio, transcripts, datasets, checkpoints, and generated WAV files under separate speaker names.
• Generates speech immediately from a single reference clip through zero-shot mode.
• Saves individual epoch checkpoints for voice-quality comparison.
• Produces multiple WAV files from structured JSON input through batch generation.
• Exposes setup, data preparation, training, and generation through both a CLI and a local web UI.
• Blocks accidental file replacement until you explicitly enable overwriting.

Use Cases

Record 15 minutes of narration in a quiet room, run vcl run, proofread the auto-generated transcript, and train a checkpoint overnight on an RTX 4090. The next morning, vcl generate produces intro and outro lines in that voice without booking a new recording session.

Paste a YouTube link to a voice actor’s earlier work, and build a training dataset directly from that audio. You no longer need to re-record lines for a game character.

Script narration for a course by writing a JSON file of line IDs and text, then run vcl generate --spec lines.json --outdir out/ to produce every audio file in a single batch pass.

Voice Clone Lab System Requirements

Install Voice Clone Lab

Clone the repository and enter its directory:

git clone https://github.com/tetsuo-ai/voice_clone_lab.git
cd voice_clone_lab

Install the core package and Gradio web interface:

pip install -e '.[ui]'

Download the patched Qwen3-TTS dependency and the 1.7B base model:

vcl setup --download

Check the GPU, dependencies, and model files:

vcl check

Train a Personal Voice

Prepare the Recording

Record 5–15 minutes of clean speech in a quiet room. Use one microphone, one speaker, and a natural reading pace. Music, overlapping speakers, room echo, and changing microphone distance weaken the training material.

Build the Speaker Dataset

Assign a name to the voice and pass the recording to the pipeline:

vcl run --speaker alex --input ~/recordings/alex.wav

A URL works as the input when the optional URL dependencies are installed:

vcl run --speaker alex --input "https://www.youtube.com/watch?v=..."

The command extracts the audio, cleans it, splits it into clips, transcribes the speech, and writes the dataset under data/voices/alex/.

Only use remote recordings that feature your own voice or a speaker who supplied explicit permission.

Correct the Transcript

Create an editable transcript review file:

vcl transcribe --speaker alex --review

Edit this file:

data/voices/alex/transcripts/transcripts_review.tsv

Apply the corrected transcript:

vcl transcribe --speaker alex --apply-review --force

Names, abbreviations, technical terms, numbers, and misheard words deserve close attention. Transcript errors enter the training data and often appear later as unstable pronunciation.

Prepare and Train the Model

Add the required Qwen audio codes:

vcl prepare --speaker alex

Start training:

vcl train --speaker alex

Use a smaller batch size on a 24 GB card:

vcl train --speaker alex --batch-size 2

Generate Speech

Generate a WAV file from the newest checkpoint:

vcl generate \
  --speaker alex \
  --text "Hello, this is my cloned voice."

Generated files appear under:

outputs/generated/alex/

Open the Web UI

Launch the local interface:

vcl ui

Open http://127.0.0.1:7860 in a browser.

The web UI contains four main tabs:

• Setup: Checks the system and model installation.
• Data: Imports recordings or URLs, runs the data pipeline, and edits transcripts.
• Train: Configures training parameters and displays the live log.
• Generate: Selects a speaker, checkpoint, and generation settings for playback.

Zero-shot mode generates speech from one reference recording and skips the training process. This mode provides a fast compatibility check before preparing a full dataset. The trained workflow creates persistent checkpoints for repeated generation.

vcl generate \
  --zeroshot \
  --ref clip.wav \
  --text "This sample uses zero-shot voice cloning."

Batch Speech Generation

Voice Clone Lab accepts a JSON file containing multiple identifiers and text lines:

[
  {
    "id": "intro",
    "text": "Welcome to the first chapter."
  },
  {
    "id": "chapter-one",
    "text": "The story begins on a quiet morning."
  }
]

Run the batch:

vcl generate --spec lines.json --outdir generated-lines/

Clone Lab Compared With Other Local Options

Pros

• Local training and generation
• Reusable speaker checkpoints
• Editable transcript review
• CLI and web UI
• Zero-shot testing mode
• Batch WAV generation

Cons

• Linux-only
• NVIDIA CUDA dependency
• 24 GB training requirement
• Manual transcript correction

Alternatives & Related Resources

• dots.tts: Free Open-Source Voice Cloning With 24-Language Support
• Free CPU-Based Text-to-Speech Tool with Voice Cloning – Pocket TTS
• 7 Best Free AI Voice Cloning Tools
• Discover More Free AI Voice Cloning Tools

FAQs

Q: Does Voice Clone Lab require model training?
A: No. Zero-shot mode generates speech directly from a reference clip. Model training creates a reusable speaker checkpoint from a larger collection of corrected recordings.

Q: Which languages does Voice Clone Lab support?
A: The Qwen3-TTS 1.7B Base model supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.

Q: Does Voice Clone Lab manage several voices?
A: Yes. Each speaker name receives separate audio, transcript, dataset, checkpoint, and generated-output directories. Use a unique --speaker value for each authorized voice.

Q: Does Voice Clone Lab require an internet connection after installation?
A: No. The tool operates offline once the base model weights and dependencies are downloaded.