Latent Vector Synthesis is a sound synthesis framework combining latent audio spaces and vector synthesis techniques.
This prototype of a Latent Vector Synthesizer incorporates a Variational Autoencoder (VAE) model trained on short single cycle waveforms that enables interpolations and explorations of sonic textures. The generated waveforms are used as part of a vector- and wavetable synthesis engine developed in Pure Data.
This project builds on the work of Tatar et al. [1, 2].
1 - Download and install Anaconda for your operating system: https://docs.anaconda.com/free/anaconda/install/index.html
2 - Open a terminal and create a new Python environment (here named 'lvs'):
conda create --name lvs python=3.10
3 - Activate your environment:
conda activate lvs
4 - Install PyTorch using conda for your operating system: https://pytorch.org
5 - Install the following Python libraries:
- Librosa
pip install librosa
- Python-osc
pip install python-osc
Install Pure Data (Pd-vanilla): https://puredata.info
1 - Open a terminal and navigate to the cloned repository.
2 - Run the Python script:
python main.py
Wait until the osc-infoformation appears (for sending/receiving osc-messages).
3 - Run latent-vector-synth.pd in Pure Data.
- Make sure the right audio output device is selected (Go to Media —> Audio Settings…)
- Press RESET and then RANDOMIZE ALL. Make sure the DSP toggle is on.
- Make sure to toggle AUDIO OUT and raise the gain.
4 - Happy droning!
[1] Kıvanç Tatar, Daniel Bisig, and Philippe Pasquier. Latent Timbre Synthesis: Audio-based variational auto-encoders for music composition and sound design applications. Neural Computing and Applications, 33(1):67–84, 2021. URL: https://link.springer.com/10.1007/s00521-020-05424-2, doi:10.1007/s00521-020-05424-2
[2] Kıvanç Tatar, Kelsey Cotton, and Daniel Bisig. Sound design strategies for latent audio space explorations using deep learning architectures. In Proceedings of Sound and Music Computing 2023, 2023.