Noisy SHD: Spiking Heidelberg Digits with Sinusoidal and Natural Noise

Noisy SHD is a collection of noisy variants of the Spiking Heidelberg Digits (SHD) dataset. The dataset contains SHD train/test spike-train files generated from source audio after adding controlled noise before conversion to HDF5. Spike conversion was performed with a locally modified version of LAUSCHER, using 700 cochlear channels and a 30 ms onset/offset window; the release README and associated code describe the conversion pipeline.

The release includes additive sinusoidal-noise conditions with fixed phase, random phase, and variable parameters, as well as natural background-noise overlay conditions derived from recorded noise sources. Each condition contains separate train.h5 and test.h5 files. The accompanying MANIFEST.csv records the split, noise family, condition parameters, estimated condition-level SNR in dB where available, file size, SHA-256 checksum, and release path.

Clean spoken-digit audio is from the Heidelberg Spiking Datasets SHD/SSC source audio. Natural-noise recordings used for the natural_noise conditions are from the Kaggle dataset “Noise Data Set” by Abdelrhaman Fakhry: https://www.kaggle.com/datasets/abdelrhamanfakhry/noise-data-set/

Documentation, release metadata, and companion code are available at: https://github.com/abdalalkilani/Noisy-SHD