sound-similarity sample browser via CLAP embeddings cosine distance
requires ~4.2 GB VRAM to run (you will need a GPU)
Download this repo, unzip it into a folder. Make sure python is installed on your system. I used python 3.10. Then:
to run after first install, run run.bat (win) or ./run.sh (mac).
run install.bat. you will need CUDA installed.
navigate to the folder, open it in terminal, and run: chmod +x install.sh then ./install.sh
create a venv python -m venv venv, enter it source venv/Scripts/activate, run pip install flask laion_clap librosa numpy torch, make sure the CUDA version of torch is installed pip install -U torch==2.4.0+cu121 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121, then run python sound-similarity-browser.py and go to http://localhost:5000/
Paste a local filepath into the Cache Management input and press Process Folder. When complete, upload a sound (use upload button or drag-drop an audio file) or type a sound description and press search.
