botfx/ringdown-damping-signals

Ring-Down Damping Signals: 12K Labelled Decay Waveforms

How this dataset was created

This is original data created programmatically — it was not collected, recorded, scraped, or derived from any external source. Each of the 12,000 signals was generated from scratch by a deterministic, seeded Python program:

1. Draw the label theta uniformly at random from [1, 5], plus a random overall base-decay rate.
2. Pick a random number of tones ("modes", 30–55), each with a random frequency, amplitude and phase from fixed ranges.
3. Set each tone's decay rate from a frequency-dependent damping law governed by theta (higher frequencies decay faster; theta controls how much faster).
4. Sum the decaying tones over a 0.5-second window (2,048 samples at 4,096 Hz), apply a random onset, add Gaussian noise at a random signal-to-noise ratio, and RMS-normalise the waveform to unit power (so loudness carries no information about theta).

Every step is seeded, so the dataset is exactly reproducible.

Contents

• train.csv — 12,000 rows; columns id, theta, s0 … s2047. One labelled training signal per row: theta is the damping-law exponent (in [1, 5]) and s0…s2047 are the 2,048 waveform samples.

Task

Recover the hidden scalar theta (the exponent of a frequency-dependent damping law) from a raw signal. Because many tones overlap under noise in a short window, this is a deliberately ill-posed inverse problem — a model trained on the signals recovers theta far better than any hand-written signal-processing rule.

License

CC0 1.0 (public domain dedication) — commercial use and redistribution permitted, no attribution required.