Frogs Exploit Statistical Regularities in Noisy Acoustic Scenes to Solve Cocktail-Party-like Problems

Abstract

Noise is a ubiquitous source of errors in all forms of communication [1]. Noise-induced errors in speech communication, for example, make it difficult for humans to converse in noisy social settings, a challenge aptly named the "cocktail party problem" [2]. Many nonhuman animals also communicate acoustically in noisy social groups and thus face biologically analogous problems [3]. However, we know little about how the perceptual systems of receivers are evolutionarily adapted to avoid the costs of noise-induced errors in communication. In this study of Cope's gray treefrog (Hyla chrysoscelis; Hylidae), we investigated whether receivers exploit a potential statistical regularity present in noisy acoustic scenes to reduce errors in signal recognition and discrimination. We developed an anatomical/physiological model of the peripheral auditory system to show that temporal correlation in amplitude fluctuations across the frequency spectrum ("comodulation") [4-6] is a feature of the noise generated by large breeding choruses of sexually advertising males. In four psychophysical experiments, we investigated whether females exploit comodulation in background noise to mitigate noise-induced errors in evolutionarily critical mate-choice decisions. Subjects experienced fewer errors in recognizing conspecific calls and in selecting the calls of high-quality mates in the presence of simulated chorus noise that was comodulated. These data show unequivocally, and for the first time, that exploiting statistical regularities present in noisy acoustic scenes is an important biological strategy for solving cocktail-party-like problems in nonhuman animal communication.

Keywords: acoustic communication; auditory masking; auditory scene analysis; cocktail party problem; comodulation masking release; mate choice; natural scene statistics; noise; sexual selection; species recognition.

Figure 1. Depictions of natural and synthetic signals and noise

(A) Spectrograms of a natural advertisement call (left, Audio Clip S1), the synthetic standard call (center), and one of several different synthetic alternative calls (right) used in this study. The depiction of a natural call illustrates two acoustic properties mimicked by synthetic stimulus calls: this signal’s pulsatile structure and bimodal frequency spectrum, with spectral peaks near 1.3 and 2.6 kHz. The synthetic standard call was used as a stimulus in Experiments 1–4. The alternative call depicted here is was used in two-alternative choice tests in Experiment 3 and differs from the standard call (50 pulses/s) in having a slower pulse rate (20 pulses/s). Oscillograms of standard and alternative calls used to create differences in call effort in Experiment 4 are illustrated in Figure S1A. Also shown here (A, far right) is a photograph of a calling male of Cope’s gray treefrog, courtesy J. C. Tanner. (B) An illustrative spectrogram (left) and the mean modulation power spectrum (right) of the natural noise generated by choruses of Cope’s gray treefrog (Audio Clip S2). The spectrogram of the natural chorus illustrates the two spectral bands of background noise in choruses arising from the mixture of vocalizations produced by calling males. The mean modulation power spectrum [20] illustrates the prominence of temporal fluctuations in amplitude (x-axis) occurring at slow rates (e.g., < 5–10 Hz). The mean depicted here was determined from an ensemble of 26 chorus recordings (see Supplemental Experimental Procedures). Each recording was truncated into 90 1-sec segments and a Gaussian spectrogram (Gaussian window bandwidth: 32 Hz, window size: 1316) was computed for each segment. A 2D FFT was computed for each Gaussian spectrogram and real values were averaged across all segments from all recordings to give an average modulation power spectrum. (C–E) Spectrograms (left) and modulation power spectra (right) of the three artificial chorus-shaped noises used in Experiments 1–4: (C) the comodulated noise (Audio Clip S3), (D) the uncorrelated noise (Audio Clip S4), and (E) the unmodulated noise (Audio Clip S5). During each behavioral test of a subject, a specified noise was broadcast continuously to simulate the ambient background noise of a chorus while one or more specified signals was broadcast periodically to simulate individual calling males. Additional details on the speaker arrangements used in behavioral experiments are provided in Figure S1B and in the Supplemental Experimental Procedures.

Figure 2. Biologically inspired analyses of chorus noise reveal significant comodulation

An anatomical/physiological model was used to determine the degree of comodulation present in natural chorus sounds. (A) The model consisted of a bank of auditory filters fitted to the audiogram of Cope’s gray treefrog. (A, left) Each filter was modeled using parameters from VIII^th nerve frequency tuning curves (FTCs) measured in previous studies of frogs (see Supplemental Experimental Procedures). An example of a previously published VIII^th nerve FTC from a frog [50] showing the rounded-exponential function (red curve) used to determine its best frequency (BF), threshold, and bandwidth 10 dB above threshold (10-dB BW). (A, center) Scatterplot showing the positive relationship between 10-dB BW and BF obtained from a meta-analysis of 1071 FTCs from seven species of frogs across 10 different published studies (see Supplemental Experimental Procedures). Units are classified as innervating either the amphibian papilla (shown in red) or the basilar papilla (shown in blue). (A, right) Diagram showing the model auditory filterbank, with the gain of each filter adjusted to the sensitivity of the midbrain audiogram (purple curve). Filters centered on the 1.3 and 2.6 kHz peaks of the advertisement call are shown in red and blue, respectively. (B) Cross-covariance analyses were conducted to quantify the magnitude of comodulation in chorus noise. (B, left) Chorus recordings (N = 26 choruses) of 1.5 min duration were filtered using the model filterbank depicted in (A, right). (B, center) Pairwise comparisons between the Hilbert envelope of the output of each frequency filter were made using cross-covariance, as illustrated here by the raw covariogram for a representative chorus recording. Below the diagonal in the raw covariogram shows peak cross-covariance magnitudes for different envelope comparisons plotted as a heat map. (B, right) The mean Z-score covariogram depicts the mean cross-covariance values, averaged across all 26 chorus recordings, as Z-scores relative to null distributions based on comparing the envelopes of frequency filter outputs across different choruses (Figure S2; Supplemental Experimental Procedures). Colors indicate the number of standard deviations beyond the mean of the null distribution. The high degree of comodulation revealed by these analyses could not be explained as merely resulting from overlap between adjacent auditory filters in the model (Supplemental Experimental Procedures).

Figure 3. Comodulated noise improves performance in several key communication tasks relative to uncorrelated noise and unmodulated noise

The three artificial chorus-shaped noises used in Experiments 1–4 where behaviorally neural and did not, by themselves, influence phonotaxis (Figure S3). Experiments 1 (A) and 2 (B) consisted of single-stimulus (no-choice) tests and revealed lower signal recognition thresholds (SRTs) in comodulated noise. In (A), points depict the proportion (± 95% exact binomial confidence intervals) of subjects responding at each of five signal-to-noise ratios (−12, −6, 0, +6, and +12 dB, or equivalent signal levels in quiet); solid lines represent fitted functions from Generalized Estimating Equations. The horizontal dashed line represents the criterion (0.5) for determining SRTs. In (B), bars depict the mean (± S.E.M.) SRTs determined using an adaptive tracking procedure. The horizontal dashed line in (B) indicates the level of performance relative to the condition with the highest threshold. Experiments 3 (C) and 4 (D) consisted of two-alternative choice tests and revealed better discrimination of sound patterns in comodulated noise. In (C), bars depict the proportion (± 95% exact binomial confidence intervals) of subjects choosing stimuli with conspecific pulse rates [P(Conspecific Pulse Rate)]. In (D), bars depict the proportions (± 95% exact binomial confidence intervals) of subjects choosing stimuli with relatively higher calling efforts [P(Higher Calling Effort)]. Horizontal dashed lines in (C) and (D) depict the level of performance expected by chance (0.5) in a two-alternative choice test.