Edit With My Voice

A local, browser-based tool for repairing spoken audio with voice-cloning assisted generation and precise audio splicing.

The goal is not to create a generic deepfake toy: it is a practical editor for replacing a wrong word or a short phrase in lectures, voiceovers, podcasts, and training material when you have authorization to use the voice.

What it does

• Loads audio/video and prepares a clean 48 kHz mono working WAV.
• Extracts reference clips from the original voice.
• Generates replacement speech locally with F5-TTS.
• Lets you define one or more timestamped edits.
• Applies replacements with time-fitting, RMS loudness matching, and crossfades.
• Exports a corrected WAV/M4A plus a before/after preview montage.

Quick start on Windows

1. Install prerequisites

• Python 3.11 from https://www.python.org/downloads/windows/ During install, enable Add python.exe to PATH.
• FFmpeg. The setup script tries to install it with winget; if that fails, install it from https://www.gyan.dev/ffmpeg/builds/ and restart the terminal.

2. Run the one-click launcher

Download or clone this repository, then double-click:

run_windows.bat

The first launch can take several minutes because it creates a virtual environment and installs AI/audio dependencies. When the browser opens, use:

http://127.0.0.1:7860

macOS / Linux

git clone https://github.com/Zer0codestuff/EditWithMyVoice.git
cd EditWithMyVoice
./scripts/setup_unix.sh
./scripts/run_unix.sh

Recommended workflow

1. Project: upload the original audio/video and click Prepare source.
2. Reference / donor clip: extract 10–30 seconds of clean voice from the same speaker.
3. Generate phrase: choose a model, paste the reference transcript, and generate the corrected phrase.
4. Edit list: set start/end timestamps for the word or phrase to replace.
5. Apply: export the corrected file and listen to the before/after preview.

For best results, generate a short phrase such as the Lightning Network, not a single isolated word. Speech sounds natural because of surrounding context, timing, breaths, and room tone.

Privacy and consent

• Use this only with the speaker's consent or with content you are authorized to edit.
• Everything runs locally except model downloads from Hugging Face during first use.
• The app does not upload your media to a hosted service.
• Generated files and uploaded media stay inside workspace/, which is ignored by Git.

Model notes

The app includes these presets:

• F5-TTS base: general default from the f5-tts package.
• Italian F5-TTS: optional preset using alien79/F5-TTS-italian from Hugging Face; useful for Italian-only audio.

Model licenses may differ from this repository's code license. Check the model card before commercial use.

Troubleshooting

ffmpeg was not found

Install FFmpeg and restart the terminal. On Windows, try:

winget install --id Gyan.FFmpeg -e

F5-TTS is slow on CPU

CPU generation works but can be slow. On NVIDIA Windows machines, install the CUDA-enabled PyTorch build in the virtual environment, then choose cuda in the UI.

The generated phrase sounds pasted in

• Use 20–40 ms crossfade.
• Generate a phrase, not just a word.
• Match the replacement to the same emotion and pace.
• Try a real donor clip from elsewhere in the same recording before using AI.

Development

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
python -m pip install -e .[dev]
pytest
ruff check .

License

MIT for the application code. Third-party models and dependencies keep their own licenses.