The soundscape dynamics of human agglomeration

We report a statistical analysis about people agglomeration soundscape. Specifically, we investigate the normalized sound amplitudes and intensities that emerge from people collective meetings. Our findings support the existence of nontrivial dynamics characterized by heavy tail distributions in the sound amplitudes, long-range correlations in the sound intensity and non-exponential distributions in the return interval distributions. Additionally, motivated by the time-dependent behavior present in the volatility/variance series, we compare the observational data with those obtained from a minimalist autoregressive stochastic model, a GARCH process, finding a good agreement.

💡 Research Summary

This paper investigates the statistical properties of acoustic recordings obtained from human gatherings, specifically focusing on the normalized sound amplitude and intensity that arise when people are congregated in a common space such as a university cafeteria. The authors recorded 16 ten‑minute sessions with a high‑fidelity condenser microphone (44.1 kHz sampling) during periods when roughly 100–200 individuals were present, and they also analyzed ten publicly available recordings from an online sound database for comparison. After subtracting the mean and dividing by the standard deviation, the raw audio signal was transformed into a normalized amplitude series A(t) and its squared counterpart (intensity) A²(t).

The probability density function (PDF) of A(t) exhibits a clear heavy‑tailed behavior: while the central region aligns well with a Gaussian distribution, values beyond four standard deviations occur far more frequently than a Gaussian would predict. This reflects the presence of extreme acoustic events such as sudden shouts, laughter, or applause that punctuate ordinary conversation.

To assess temporal correlations, the authors applied detrended fluctuation analysis (DFA) to the intensity series. The fluctuation function F(n) scales as n^h with a Hurst exponent h ≈ 0.88, indicating strong long‑range memory (h > 0.5). This exponent is remarkably consistent across all experimental recordings (average h = 0.88 ± 0.01) and the external web recordings (average h = 0.89 ± 0.01).

The study further examines return intervals τ, defined as the time between successive instances when the intensity exceeds a chosen threshold q (e.g., q = 1, 2, 5). For uncorrelated Gaussian noise, τ follows an exponential law, but the empirical τ distributions are better described by stretched‑exponential or Weibull forms. By normalizing τ with its mean, the authors find a scaling exponent γ ≈ 0.24, which satisfies the theoretical relation γ = 2(1 − h). This demonstrates that the long‑range correlations affect the statistics of extreme‑event waiting times.

Volatility, measured as the local standard deviation of A(t) over sliding windows ranging from 0.01 s to 1 s, displays a power‑law tail p(v) ∝ v^−η with η ≈ 4.3 (experimental data) and η ≈ 4.9 (web data). The heavy‑tailed volatility indicates non‑stationary fluctuations and the possibility of large, abrupt changes in acoustic energy.

Motivated by the observed volatility clustering and memory, the authors model the series using a financial‑style GARCH(1,1) process: x_t = σ_t ξ_t, σ_t² = α₀ + α₁ x_{t−1}² + β₁ σ_{t−1}², with ξ_t drawn from a standard normal distribution. By enforcing unit variance (σ_x² = 1), the three parameters reduce to two independent ones. Least‑squares fitting yields α₁ = 0.011, β₁ = 0.9889, and consequently α₀ = 0.001. Simulated series from this GARCH model reproduce the empirical PDF of amplitudes, the DFA exponent h, the return‑interval distribution, and the volatility power‑law tail with high fidelity. The near‑unit sum α₁ + β₁ ≈ 0.9999 generates an effective characteristic time τ_c ≈ 10⁴ s, allowing the model to mimic the observed long‑range decay despite its underlying exponential autocorrelation.

In summary, the acoustic environment of human agglomerations exhibits four key statistical signatures: (i) non‑Gaussian heavy‑tailed amplitude distribution, (ii) long‑range correlated intensity, (iii) non‑exponential return‑interval statistics, and (iv) volatility with a power‑law tail. A minimalist GARCH(1,1) framework captures all these features, suggesting that the collective behavior of people in a shared space can be described by mechanisms akin to those governing financial markets. The authors propose that further work should incorporate more detailed measurements and multi‑scale interaction models to deepen the understanding of social acoustic dynamics.