Training Input Configuration Dataset for SA-2A Compressor Modeling

This dataset was created for the study: “Comparing Training Input Configurations for Temporal Convolutional Network-Based Modeling of an Analog Optical Compressor” (unpublished). It is designed for training and evaluating deep learning models for analog audio compressor emulation, with a particular focus on the influence of DAC–ADC conversion in the input signal path. The dataset consists of paired recordings organized under two input conditions: 1) Original digital input (Model A input)2) DAC–ADC loopback processed input (Model B input) For both conditions, the target signal is the same analog optical compressor output recorded from a Stam Audio SA-2A (Gen1). The outboard output is shared between Model A and Model B and serves as the common supervised-learning target. Dataset contents include: - 15 min Original Input from SignalTrain (Model A input)- 20 min Original Input from SignalTrain (Model A input)- 15 min DAC–ADC Loopback Input (Model B input)- 20 min DAC–ADC Loopback Input (Model B input)- 15 min Outboard Output (shared target)- 20 min Outboard Output (shared target) All recordings were performed at 44.1 kHz, 32-bit floating-point, in mono. The dataset is sample-aligned using a cross-correlation-based latency estimation procedure implemented with Audio-Loopback-Recorder (https://github.com/JYKlabs/Audio-Loopback-Recorder) to ensure temporal consistency between input and output pairs. All recordings were conducted with the Gain parameter fixed at 20. The Peak Reduction parameter was varied from 0 to 100 in increments of 5, resulting in 21 conditions per dataset. Each folder contains 21 audio files corresponding to these Peak Reduction settings. File names follow a consistent convention: - Model A input: `SA2A_input_gain20_peakX.wav`- Model B input: `directardinput_gain20_peakX.wav`- Output (shared target): `outboardinput_gain20_peakX.wav` where `X` denotes the Peak Reduction value (0–100 in steps of 5). The original input audio is derived from the publicly available SignalTrain LA-2A dataset:https://doi.org/10.5281/zenodo.3824876 If you use this dataset, please cite both this repository and the associated study (once published).