The encoding of individual identity in dolphin signature whistles: how much information is needed?

Abstract

Bottlenose dolphins (Tursiops truncatus) produce many vocalisations, including whistles that are unique to the individual producing them. Such "signature whistles" play a role in individual recognition and maintaining group integrity. Previous work has shown that humans can successfully group the spectrographic representations of signature whistles according to the individual dolphins that produced them. However, attempts at using mathematical algorithms to perform a similar task have been less successful. A greater understanding of the encoding of identity information in signature whistles is important for assessing similarity of whistles and thus social influences on the development of these learned calls. We re-examined 400 signature whistles from 20 individual dolphins used in a previous study, and tested the performance of new mathematical algorithms. We compared the measure used in the original study (correlation matrix of evenly sampled frequency measurements) to one used in several previous studies (similarity matrix of time-warped whistles), and to a new algorithm based on the Parsons code, used in music retrieval databases. The Parsons code records the direction of frequency change at each time step, and is effective at capturing human perception of music. We analysed similarity matrices from each of these three techniques, as well as a random control, by unsupervised clustering using three separate techniques: k-means clustering, hierarchical clustering, and an adaptive resonance theory neural network. For each of the three clustering techniques, a seven-level Parsons algorithm provided better clustering than the correlation and dynamic time warping algorithms, and was closer to the near-perfect visual categorisations of human judges. Thus, the Parsons code captures much of the individual identity information present in signature whistles, and may prove useful in studies requiring quantification of whistle similarity.

Figure 1. Examples of the spline-smoothed whistles (blue line), on top of manually extracted curves (red points).

Figure 2. An example of the dynamic time-warping matching of two whistle profiles.

The left frame shows the original signals on arbitrary time and frequency axes. The right frame shows the red sample having undergone a dynamic time-warping transformation to produce the minimum least-squares distance from the blue sample. Note how the spacing of the points in the curve have been varied.

Figure 3. Sensitivity of the algorithm performance (normalised mutual information) as the number of Parsons segments is varied.

Figure 4. Sensitivity of the algorithm performance (normalised mutual information) for all metrics (Parsons, DTW, correlation, and random control), and all clustering algorithms (ART, k-means, and Hierarchical), as the number of clusters is varied.

Figure 5. Normalised mutual information for different n-Parsons encodings, with each of the clustering algorithms.

Error bars indicate the standard error of the 100 bootstrapped iterations.

Figure 6. Boxplots showing the normalised mutual information (NMI) for the different encodings, compared to human clustering (A).

The panels show ART clustering (left), k-means clustering (middle) and hierarchical clustering (right), using the different proximity metrics. Each box shows the 25^th and 75^th centiles, with the median indicated as a red line. Whiskers show the extreme values (±2.7σ) using the Matlab boxplot function. Letters indicate Tukey HSD post-hoc tests at the p<0.05 level, following ANOVA.