Source-linked AI summary
Estimation of Room Impulse Responses from Handclaps
Shih-Yu Lai, Kyung Yun Lee, Nils Meyer-Kahlen, Eloi Moliner, Bing-Yu Chen, Vesa Välimäki
TL;DR
Unknown, variable handclap excitations make direct RIR estimation challenging. The paper introduces an anechoic handclap dataset and trains a supervised neural regressor to estimate RIRs without knowing the excitation. The regressor outperforms windowing-based baselines in synthetic evaluation, and real-room estimates are consistent across handclaps at the same location.
Problem
Handclap excitation spectra vary substantially, making it challenging to estimate an RIR from a recording that entangles unknown excitation and room response.
Source: §1 · p. 1Method
The authors use an anechoic handclap dataset to train a neural regressor on synthetic paired examples to estimate RIRs without a known excitation.
Source: §1 · p. 1Results
The neural regressor improves on windowed-excitation baselines across all reported metrics for measured and simulated RIRs, while a gap remains against known-excitation deconvolution.
Source: §4.4 · pp. 3–4Takeaways & Limitations
Real-world tests show consistent RIR estimates across different handclap recordings made in the same room.
Source: §5 · p. 4Training on measured RIRs includes measurement-hardware transfer functions, so the network regresses toward average loudspeaker and microphone responses rather than pure room acoustics.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Handclaps provide an equipment-free excitation for room acoustics, but their unknown and variable source waveform makes room impulse response (RIR) estimation challenging. In this work, we investigate whether RIRs can be estimated directly from handclaps. To this end, we introduce an anechoic handclap dataset containing 2,540 claps from 17 participants, designed to capture variability across natural claps and different hand configurations. We first establish the performance attainable when the excitation clap is known using regularized deconvolution, and show that approximating the unknown excitation by windowing the direct sound from the reverberant recording is insufficient. To estimate the RIR without a known excitation, we propose using the anechoic handclap recordings to train a deep neural network with a supervised regression objective. Evaluated on a controlled synthetic benchmark, the proposed neural regressor significantly outperforms windowing-based baselines across all instrumental metrics. Furthermore, we test the proposed method on handclap recordings measured in real acoustic spaces, showing that the inferred RIR spectra are consistent across different handclap measurements taken in the same room location. These results showcase the feasibility of directly estimating RIRs from natural handclaps without requiring knowledge of the excitation signal.
1. INTRODUCTION
The paper tackles RIR estimation from a single handclap when the excitation is unknown and variable, using an anechoic dataset to train a neural regressor that outperforms windowing-based baselines on a controlled synthetic benchmark.
1. INTRODUCTION
- Prior handclap work estimates reverberation time and energetic room-acoustic parameters, and has also approximated binaural RIRs.Handclaps offer an equipment-free, impulsive excitation with broad-frequency energy, but their spectra vary across people, postures, and repeated claps [3–13].Source: §1 · p. 1
- Estimating an RIR from one recorded handclap is difficult because the unknown excitation and unknown room response are entangled in the recording.Blind deconvolution is particularly ill-posed given variation across handclap executions and participants.Source: §1 · p. 1
- Windowing the direct path to approximate the excitation is insufficient for reliable RIR recovery.Source: §1 · p. 1
- The proposed neural regressor, trained using a new anechoic handclap dataset, significantly outperforms compared baselines across acoustic, spectral, and waveform error metrics on a controlled synthetic benchmark.The dataset captures excitation variability and enables construction of training pairs with known RIRs.Source: §1 · p. 1
2. ANECHOIC HANDCLAP DATASET
The dataset comprises isolated anechoic handclaps recorded from 17 participants, with a protocol spanning multiple hand configurations and a final collection of 2,540 eight-channel claps.
2. ANECHOIC HANDCLAP DATASET
- Seventeen participants clapped at the center of an eight-microphone circular array with a 2 m radius.The protocol requested 15 claps for each hand posture and 150 claps per participant, though actual counts varied.Source: §2 · p. 2
- The publicly available dataset contains 2,540 eight-channel anechoic handclaps, segmented semi-automatically and manually checked for spurious peaks.Extraction windows ran from 10 ms before to 50 ms after each peak.Source: §2 · p. 2
- Handclap spectra vary substantially between individual claps, while average spectra differ across requested handclap types.All handclaps show bandpass characteristics; center frequency depends to some extent on handclap type.Source: §2 · p. 2
- Across the requested postures, median clap durations ranged from 4.85 ms for A3 to 6.46 ms for cupped-hand A1+.The examples also show sharp high-frequency spectral minima.Source: §2 · p. 2
3. METHODS
The methods formalize recovery of an RIR from a single handclap with unknown excitation, then compare known-excitation deconvolution, windowed-excitation baselines, and supervised neural regression.
3.1. Identifiability and known-excitation reference
- Known-excitation deconvolution becomes unstable near spectral minima, so Tikhonov regularization trades noise amplification against bias.At bins where the excitation magnitude is small relative to the regularization parameter, the estimate is driven toward zero.Source: §3.1 · p. 2
3.2. Windowed-excitation baseline
- The windowed baseline substitutes the measurement’s initial direct-sound segment for the unknown excitation, using 3-ms or 6-ms windows.A 6-ms window can include early reflections, while a 3-ms window truncates the excitation.Source: §3.2 · pp. 2–3
3.3. Neural regressor
- The neural regressor maps a reverberant handclap directly to an RIR, trained on synthetic pairs made by convolving anechoic claps with measured RIRs.This supervised approach requires no estimate of the excitation waveform.Source: §3.3 · p. 3
- The network adapts a two-stage time–frequency/time-domain architecture to single-channel input and combines waveform MSE with magnitude-compressed STFT loss.Source: §3.3 · p. 3
4. EXPERIMENTS AND RESULTS
The experiments evaluate supervised RIR regression on controlled synthetic data and phone-recorded handclaps, comparing it with windowed-excitation and known-excitation approaches. The neural regressor outperforms windowed-excitation baselines across reported metrics, and its estimates show lower spectral dispersion across handclaps in the same room.
4.1. Training and testing data setup
- The synthetic test set contains 432 clap–RIR observations, evenly split between measured and simulated RIRs; training uses 12 participants and testing uses three held-out participants.Training uses room-disjoint splits; OpenAIR RIRs are reserved for testing, and observations are synthesized from anechoic claps convolved with RIRs.Source: §4.1 · p. 3
4.2. Model architecture and training details
- The regressor maps a 1-s reverberant handclap to a 1-s RIR using an STFT-based convolutional NCSN++ stage and a waveform U-Net that refines early reflections.The U-Net output is added to the coarse estimate from the first stage.Source: §4.2 · p. 3
4.3. Evaluation metrics
- Evaluation reports broadband EDC, absolute EDT error, log-spectral error, and waveform NRMSE; lower values indicate better performance.EDC measures normalized broadband decay-shape error, while LSE compares log-magnitude STFTs.Source: §4.3 · p. 3
4.4. RIR estimation evaluation
- Known-excitation Tikhonov deconvolution reduces measured-RIR NRMSE from 0.061 to 0.022, but unregularized deconvolution is slightly better for simulated RIRs (0.060 vs. 0.069).For unknown excitation, a 3-ms window truncates the clap, while a 6-ms window includes early reflections and performs worse.Source: §4.4 · pp. 3–4
- The neural regressor improves on windowed-excitation baselines across all reported metrics for both measured and simulated RIRs; on measured RIRs, LLSE falls from 0.20 to 0.112 and LNRMSE from 2.7 to 0.733.These examples are relative to the best-performing windowed baseline; a gap remains versus the known-excitation reference.Source: Table 1 · p. 4, §4.4 · p. 4
4.5. RIR estimation consistency on measured handclaps
- Across seven rooms, mean spectral dispersion is 3.43 dB for neural estimates, versus 6.45 dB for raw claps, 6.76 dB for the 3-ms baseline, and 6.12 dB for the 6-ms baseline.Fig. 3 shows estimated spectra more consistent across positions than raw handclaps, whose resonances vary with the clap.Source: §4.5 · p. 4, Figure 3 · p. 4
5. CONCLUSION
The proposed neural regressor outperforms windowed-excitation deconvolution and produces consistent RIR estimates across handclaps recorded at the same room location. The study also identifies hardware responses in measured-RIR targets and a remaining gap from known-excitation deconvolution as limitations.
5. CONCLUSION
- The neural regressor outperforms windowed-excitation deconvolution, while real-world estimates remain consistent across handclaps recorded in the same room.The real-world test indicates generalization beyond synthetic data and decoupling of RIR estimates from individual handclap spectral variation.Source: §5 · p. 4
- When the exact excitation is known, regularized deconvolution is effective because handclaps contain sufficient spectral information.This known-excitation reference still outperforms the blind method, leaving a performance gap.Source: §5 · p. 4
- Measured-RIR training targets include measurement-hardware transfer functions, so the network regresses toward average loudspeaker and microphone responses rather than pure room acoustics.The authors identify explicitly decoupling device responses from targets as future work.Source: §5 · p. 4
- Further improvement may come from using multiple handclap recordings to promote more robust RIR estimation.The study also establishes the anechoic handclap dataset to capture natural excitation variations.Source: §5 · p. 4