ai-choir transforms any singing vocal into a choir rendition, best used to add crowd-like doubles or generally as a supercharged chorus effect. It works by using AI to transform the vocal into several different voices (sopranos, tenors and altos), applying detuning/dequantizing, and layering them together.
demo: https://www.tiktok.com/@popeskasounds/video/7375255624751353131
All you need is python3 https://www.python.org/downloads/ Otherwise, all the setup will be handled for you by the script. The first time you run the script will take significantly longer as it installs all the dependencies, and will take up about 4 GB of space (the voice models are about half a gig each!)
Transform any wav file like this: python3 gen.py ./path/to/file.wav
Additionally, there's a config.json where you can play around with some parameters:
cleanup (true/false) - whether the script should cleanup iterim files
convolution_reverb_dry_wet (0 - 1.0) - adds convolution reverb to the output. switch out the impulse.wav file for whatever impulse response you want!
stereo_spread (0 to 4.0) - how panned the voices should be. 0 is mono, 4.0 will be hard left/right.
base_detune (0 t0 0.05) - how detuned the voices should be
detune_drift (0 to 25% of base_detune) - how much the voices should fluctuate around the base detuning
detune_frequency (0 to 5.0) - how long voices will linger on a detuned note
output_gain (-inf to inf) - apply gain to the output
If you want to modify the voices of the choir, you can stick any so-vits-svc models into the models directory, just make sure the .pth file is called model.pth. You may want to tinker with the panning for your model in gen_combine.py.
Would love to see additional voice models, processing, and configurations improve the end result of this project. Throw up a PR for anything that sounds cool, and I'll check it out!
ai-choir is licensed under the MIT License. You are free to use, modify, and distribute this software as permitted by the license.