Audio Separate AI is an application allowing to separate vocals from audio files.
This works using demucs from the Facebook research team, made based on the paper Hybrid Spectrogram and Waveform Source Separation.
In terms of architecture, this application uses services on AWS which are spun up with Terraform to do the following in order:
- Push the unprocessed file on an S3 bucket and push its hash in an SQS queue
- Pull the unprocessed file and process it on a lambda, with layer caching for the AI model
- Push the processed file to the S3 bucket and notify the API service
- Allow for downloading the processed vocals & sounds on the client
- Python 3.8.x to 3.10.x
- Terraform with set crendentials for publishing the app on AWS
- Create a virtual environment with
python -m venv .venv - Activate the virtual environment with
source .venv/bin/activate - Install required packages
pip install -r api-service/requirements.txt && pip install -r splitter-service/requirements.txt - Initialize terraform backend
terraform init - Apply any additional features
terraform apply
- Run the
api-servicewithpython3 api-service/main.py - Publish an audio to the page
- Download separated audios