Handclaps offer an equipment-free excitation for room acoustics, but the source waveform varies with the person and hand configuration. We study supervised estimation of a one-second room impulse response (RIR) from a reverberant handclap, using controlled training pairs made with recorded anechoic handclaps and RIRs.
This page provides additional real phone-recording examples for listening and visual inspection. The phone recordings have no paired reference RIR, so these examples illustrate model behavior and do not establish reconstruction accuracy.
iPhone 15 recordings from seven rooms, originally at 48 kHz and resampled to 44.1 kHz. Every room contains eight protocol modes with five repetitions each. The default example for each room and mode is the existing unclipped recording with the highest pre-handclap SNR, selected using only the input.
Both players use the original one-second model interval. Each WAV is independently peak-normalized to −1 dBFS for listening only; no denoising or filtering is applied. Listen for the onset and decay of each inferred response, and compare the two rooms. There is no paired reference RIR.
Download the six paper-selected P1/P2/P3 pairs (ZIP) · Recording inventory
P1 · repetition 3 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
P1 · repetition 3 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
The same input-selection rule and frozen Regression model are used for the remaining rooms. All five repetitions are available through each recording menu.
P1 · repetition 1 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
P1 · repetition 3 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
P1 · repetition 3 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
P1 · repetition 3 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
P1 · repetition 5 · input-selected example
Download recording · Download inferred RIR
Top: recorded claps. Bottom: inferred RIRs. Fixed input-selected examples, 0–0.9 s and 0.1–10 kHz, with a shared color scale.
Expand the figures below to inspect the same input-selected examples. Spectrogram power is normalized by each panel’s mean and displayed with one shared −20 to +30 dB scale.
The temporal panels show 0–500 ms RMS envelopes, with a 5 ms window and approximately 1 ms spacing. They use the original digital levels without peak normalization, with shared −70 to 0 dB limits and digital full scale equal to one for every curve. The 500 ms window replaces an earlier 250 ms crop, which ended while the reverberant hallway was still only about 19 dB into its decay.
These combined figures use 0.1–15 kHz spectra from the first 0.9 s, with 1/12-octave power averaging and the original per-signal mean FFT-bin power normalization over 0.1–10 kHz. All spectral panels share −40 to +20 dB limits. Only the temporal panels use unnormalized digital amplitudes; these are not calibrated sound-pressure levels.
Paper spectra, 15 kHz · RMS envelopes (dB) · Raw linear waveforms
Photographs from the seven phone-recording rooms.
The source handclaps were recorded with eight microphones in the large anechoic chamber at Aalto University. Below are five retained examples from different participants. These listening files preserve each complete retained event on the primary microphone channel.
Participant IDs are kept for provenance. We do not assign P1/A1 or other pose labels to these individual anechoic events, because their retained-event metadata do not establish those labels.
Block 01, clap 01 · 0.26 s
Block 01, clap 01 · 0.26 s
Block 01, clap 01 · 0.26 s
Block 03, clap 01 · 0.26 s
Block 01, clap 01 · 0.26 s
Hand-pose reference with spectra and pressure waveforms. The listening examples above are separate retained events; their individual pose labels are not established by the available metadata.
Project code and release information · Benchmark participant split
https://akinesia112.github.io/ClapRIR/Asset manifest · Source and processing notes · Website repository