Datasets for Asymptotically Stable Recurrent Neural Networks

This record contains three datasets for training and evaluating controllable neural models of guitar/bass audio effects. The data were recorded from three physical devices: the ProCo RAT, Darkglass DFZ, and Boss CS-3. The record also contains results obtained from experiments conducted in this paper: https://arxiv.org/abs/2509.15622 The input audio was taken from four one-hour-long audio files consisting of electric and bass guitar playing. The dry input audio files are included in this record as well. Bass Guitars: ESP-LTD B-5 - esp_ltd_b5.flac (used for training sets) Harley Benton MB-5 - hb_mb_5.flac (used for evaluation sets) Electric Guitars: Ibanez RGD-7 ALMS - ibanez_rgd7alms.flac (used for training sets) Ibanez RG-7321 - ibanez_rg7321.flac (used for evaluation sets) For more information on the capture process, see https://urn.fi/URN:NBN:fi:aalto-202512169338 Dataset Structure The directory structure for the captured samples is device/{eval,train}/file.wav The files contain both dry and wet audio. The dry audio is in channel one, while the wet audio is in channel two. The used control positions were encoded in the filenames using values 0 to 100. For example, the rat with all controls at noon: 50,50,50.wav For the DFZ and the CS-3, multiple samples were captured using the same control settings, and the last number is used to differentiate the different samples taken using the same settings. For example, the 32nd sample with controls at noon for the DFZ: 50,50,32.wav