Reproducibility Package for a Temporal Convolutional Network-Based Analog Compressor Modeling

This Zenodo record provides the reproducibility assets associated with the study: Comparing Training Input Configurations for Temporal Convolutional Network-Based Modeling of an Analog Optical Compressor (unpublished manuscript). This record is separate from the previously released training dataset and contains the supplementary materials required to reproduce model training, inference, objective evaluation, and perceptual listening experiments reported in the study. Contents include: 1. Best model checkpoints- 15 min Model A: epoch_146.pt, epoch_146.pth- 15 min Model B: epoch_141.pt, epoch_141.pth- 20 min Model A: epoch_150.pt, epoch_150.pth- 20 min Model B: epoch_147.pt, epoch_147.pth 2. MUSHRA listening-test audio files- 15_min_mushra_wav- 20_min_mushra_wav Each folder contains the WAV stimuli used in the listening tests, including:- hidden anchors band-limited at 3.5 kHz and 7 kHz- Model A outputs- Model B outputs- original hardware reference signals The stimuli are organized by source material (bass, drum, acoustic guitar, vocal), training duration (15 min or 20 min), and compression condition. 3. Training code and logs- 15 min model A training log.csv- 15 min model B training log.csv- 20 min model A training log.csv- 20 min model B training log.csv- Training code example.py The training script is provided as an example implementation, with filenames and save paths adapted to the experiment setup. 4. Listening-test spreadsheet- mushra_test.xlsx 5. Offline inference script- SA2AclonewithTCN.py This script can be used for offline WAV-to-WAV inference with `.pth` checkpoints. 6. Real-time plugin supportThe exported TorchScript `.pt` files can be loaded into the real-time SA-2A plugin available at:https://github.com/JYKlabs/SA-2A-TorchScript-Plugin The `.pth` checkpoint files can be loaded into the offline PyTorch inference application available at:https://github.com/JYKlabs/SA2A-clone-with-TCN-and-python 7. Revision Analysis Materials The folder "Revision Analysis Materials" contains additional materials generated during the manuscript revision process in response to reviewer comments regarding held-out test-set evaluation and objective statistical analyses. Contents include: - ALL_EPOCH_RESULTS.csv- A_15_sample_metrics.csv- B_15_sample_metrics.csv- A_20_sample_metrics.csv- B_20_sample_metrics.csv- Epochwise_Heldout_Testset_Evaluation.py- Testset_Statistical_Evaluation.py- Readme.txt These materials were generated using the same experimental settings employed in the original study, including: - clip length: 512 samples- validation split: 15%- test split: 5%- fixed split seed: 42- identical dataset index CSV files used during the original experiments. "Epochwise_Heldout_Testset_Evaluation.py" performs held-out test-set evaluation for all saved checkpoints and generates the epoch-wise objective results used to support model selection analyses reported in the revised manuscript. "Testset_Statistical_Evaluation.py" reproduces the original held-out test partition using the preserved dataset index CSV files and split seed (42), computes sample-level LSD and MRSTFT values for the selected models, and generates the data used for the paired statistical analyses reported in the revised manuscript. The CSV files (ALL_EPOCH_RESULTS.csv and *_sample_metrics.csv) contain the objective evaluation results obtained from these analyses and correspond directly to the additional results introduced during the revision process. Related dataset The aligned training dataset used to train these models is released separately at DOI: 10.5281/zenodo.19720792 Notes This Zenodo record contains research artifacts and supplementary materials rather than the training dataset itself. The "Revision Analysis Materials" folder documents the additional objective analyses introduced during peer review and supports the reproducibility of the revised manuscript.