Modeling Time-Variant Responses of Optical Compressors With Selective State Space Models

This code repository for the article _Modeling Time-Variant Responses of Optical Compression with Selective State Space Models, Journal of the Audio Engineering Society, 2025 March - Volume 73 Number 3.

This repository contains all the necessary utilities to use our architectures. Find the code located inside the "./Code" folder, and the weights of pre-trained models inside the "./Weights" folder

Visit our companion page with Audio Samples

Folder Structure

./src
├── Code
└── Weights
    ├── CL1B_analog
    │   ├── ED-CNNCL1B_analog
    │   ├── EDbaselineCL1B_analog
    │   ├── LSTMbaselineCL1B_analog
    │   ├── LSTMCL1B_analog
    │   ├── MambaCL1B_analog
    │   └── S4DCL1B_analog
    └── LA2A_analog
        ├── ED-CNNLA2A_analog
        ├── EDbaselineLA2A_analog
        ├── LSTMbaselineLA2A_analog
        ├── LSTMLA2A_analog
        ├── MambaLA2A_analog
        └── S4DLA2A_analog

Contents

1. Datasets
2. How to Train and Run Inference
3. VST Download

Datasets

Datsets are available here

Our architectures were evaluated on two optical compressors:

• Teletronix LA-2A optical compressor
• TubeTech CL 1B optical compressor

How To Train and Run Inference

First, install Python dependencies:

cd ./Code
pip install -r requirements.txt

To train models, use the starter.py script. Ensure you have loaded the dataset into the chosen datasets folder

Available options:

• --model_save_dir - Folder directory in which to store the trained models [str] (default ="./models")
• --data_dir - Folder directory in which the datasets are stored [str] (default="./datasets")
• --datasets - The names of the datasets to use (LA2A, CL1B). [ [str] ] (default=[" "] )
• --comp - The names of the device to consider (LA2A, CL1B). [ [str] ] (default=[" "] )
• --epochs - Number of training epochs. [int] (default =60)
• --model - The name of the model to train ('LSTM', 'ED', 'LRU', 'S4D', 'S6') [str] (default=" ")
• --batch_size - The size of each batch [int] (default=8 )
• --units = The hidden layer size (amount of units) of the network. [ [int] ] (default=8)
• --mini_batch_size - The mini batch size [int] (default=2048)
• --learning_rate - the initial learning rate [float] (default=3e-4)
• --only_inference - When True, skips training and runs only inference on the pre-model. When False, runs training and inference on the trained model. [bool] (default=False)

Example training case:

cd ./Code/

python starter.py --datasets LA2A --comp LA2A --model LSTM --epochs 500 

To only run inference on an existing pre-trained model, use the "only_inference". In this case, ensure you have the existing model and dataset (to use for inference) both in their respective directories with corresponding names.

Example inference case:

cd ./Code/
python starter.py --datasets LA2A --comp LA2A --model LSTM --only_inference True

VST Download

VSTs

Bibtex

If you use the code included in this repository or any part of it, please acknowledge its authors by adding a reference to these publications:

@article{simionato2025modeling,
  title={Modeling Time-Variant Responses of Optical Compressors with Selective State Space Models},
  author={Simionato, Riccardo and Fasciani, Stefano},
  journal={Journal of Audio Engineering Society},
  volume={73},
  number={3},
  pages={144–165},
  year={2025}
}