The Concatenator

This codebase has a basic python prototype of "The Concatenator," a fast, real time capable implementation of concatenative synthesis.

Installation

Installing Basic Requirements

If you do not have python on your system, follow the instructions at this link to download and install miniconda. Then, open up the anaconda prompt as indicated at that link, and change directories into TheConcatenator folder.

Then, regardless of whether this is a fresh install or an old install, make sure you have the following dependencies installed: numpy, matplotlib, scikit-learn, numba, librosa

pip install numpy matplotlib scikit-learn numba librosa

It is also highly recommended that you install cuda and torch if you have a cuda-capable Nvidia GPU. As of the writing of this readme (4/29/2024), the commands to install the best versions of this in miniconda are the following, in sequence:

conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

Beware that this will download ~10GB of data to your computer, and so it may take a moment.

Installation for real time

If you want to do real time, be sure that portaudio is installed. On mac, for instance

brew install portaudio

Then, you can install pyaudio

pip install pyaudio

Example Usage

Type

python musaic.py --help

for all options. For instance, to create the "Let It Bee" example offline, you can run

python musaic.py --corpus corpus/Bees_Buzzing.mp3 --target target/Beatles_LetItBe.mp3 --minFreq 0 --maxFreq 8000 --particles 1000 --pd 0.95 --temperature 50 --p 5 --device np --result 1000Particles.wav

The above works best on a mac. But if you're on windows and linux and you have pytorch installed with cuda support, you can run the following instead, which will go much faster by using the GPU

python musaic.py --corpus corpus/Bees_Buzzing.mp3 --target target/Beatles_LetItBe.mp3 --minFreq 0 --maxFreq 8000 --particles 1000 --pd 0.95 --temperature 50 --p 5 --device cuda --result 1000Particles.wav

You can also use torch with the cpu, which sometimes threads better than numpy

python musaic.py --corpus corpus/Bees_Buzzing.mp3 --target target/Beatles_LetItBe.mp3 --minFreq 0 --maxFreq 8000 --particles 1000 --pd 0.95 --temperature 50 --result 1000Particles.wav --p 5 --device cpu

Then, based on whichever device is working the fastest, you can launch a real time session by changing --target to "mic". It also helps to switch to mono instead of stereo so the code goes faster. For instance, suppose cuda works for you. Then you can say

python musaic.py --corpus corpus/Bees_Buzzing.mp3 --target mic --minFreq 0 --maxFreq 8000 --particles 1000 --pd 0.95 --temperature 50 --p 5 --device cuda --stereo 0 --result 1000Particles.wav

Depending on your system, you may be able to use more particles and more activations (the --p parameter). But if you get "underrun occurred" regularly in the console, it means that the program can't keep up with the parameters you've chosen in the real time scenario.