Overview
This repository extends Facebook Research's repository by adding the capability to visualize the results (shown below) via video_generator.py
and video_generator.ipynb
. The Jupyter notebook is also available in Colab, which benefits greatly from GPU acceleration. This visualization is an .mp4 video file which also includes the original audio. The audio file can be broken into several "sets", generating a video file for each set.
Installation
In order to read MP3 or MP4 files ffmpeg is required. To install it on Windows 10 follow these steps:
- Download the full version from (here)[https://www.gyan.dev/ffmpeg/builds/]
- Extract the Zip file and paste it at
C:\
or wherever you prefer - Add the filepath
[ffmeg root dir]/bin
to thePATH
environment variable - Restart computer
To create a conda environment with Demucs, use: conda env create --file=environment-cpu.yml
or conda env create --file=environment-cuda.yml
. Also, requirements_minimal.txt
has been provided for pip install
.
Desired Upgrades
- Use multi-processing to generate video frames with a parallel for loop.
- Diarize vocals into individual singers.