Distributed acoustic cues for caller identity in macaque vocalization

Abstract

Individual primates can be identified by the sound of their voice. Macaques have demonstrated an ability to discern conspecific identity from a harmonically structured ‘coo’ call. Voice recognition presumably requires the integrated perception of multiple acoustic features. However, it is unclear how this is achieved, given considerable variability across utterances. Specifically, the extent to which information about caller identity is distributed across multiple features remains elusive. We examined these issues by recording and analysing a large sample of calls from eight macaques. Single acoustic features, including fundamental frequency, duration and Weiner entropy, were informative but unreliable for the statistical classification of caller identity. A combination of multiple features, however, allowed for highly accurate caller identification. A regularized classifier that learned to identify callers from the modulation power spectrum of calls found that specific regions of spectral–temporal modulation were informative for caller identification. These ranges are related to acoustic features such as the call’s fundamental frequency and FM sweep direction. We further found that the low-frequency spectrotemporal modulation component contained an indexical cue of the caller body size. Thus, cues for caller identity are distributed across identifiable spectrotemporal components corresponding to laryngeal and supralaryngeal components of vocalizations, and the integration of those cues can enable highly reliable caller identification. Our results demonstrate a clear acoustic basis by which individual macaque vocalizations can be recognized.

Finding, visualizing, and quantifying latent structure across diverse animal vocal repertoires

“…Caller identity is evident in latent projections of all four datasets ( Fig 4 ). The first dataset is comprised of macaque coo calls, where identity information is thought to be distributed across multiple features including fundamental frequency, duration, and Weiner entropy [ 27 ]. Indeed, the latent projection of coo calls clustered tightly by individual identity (silhouette score = 0.378; Fig 4A ).…”

“…Egyptian fruit bat pup isolation calls, which in other bat species are discriminable by adult females [ 45 , 45 , 46 ] clearly show regions of UMAP space densely occupied by single individual’s vocalizations, but no clear clusters (silhouette score = -0.078; Fig 4C ). In the marmoset phee call dataset [ 47 ] it is perhaps interesting that given the range of potential features thought to carry individual identity [ 27 ], phee calls appear to lie along a single continuum where each individual’s calls occupy overlapping regions of the continuum (silhouette score = -0.062; Fig 4D ). The silhouette score for each species was well above chance (H(2) > 20, p < 10 -5 ).…”

“…In machine learning parlance, the process of determining the relevant features, or dimensions, of a particular dataset, is called feature engineering . Engineered features are ideal for many analyses because they are human-interpretable in models that describe the relative contribution of those features as explanatory variables, for example explaining the contribution of the fundamental frequency of a coo call in predicting caller identity in macaques [ 27 ]. As with other human-centric heuristics, however, feature engineering has two caveats.…”

Optimal features for auditory categorization

“…Conceptually, our MIFs may be similar to ‘image signatures’ obtained by recently developed unsupervised methods 32 (see Supplementary Discussion). Our approach is complementary to alternative experimental approaches, such the characterization of neural tuning along an exhaustive list of call parameters 33 , characterizing call tuning as tuning for regions of the modulation spectrum 34 – 36 , and combinations of these methods in conjunction with machine learning tools 37 (see Supplementary Discussion). Our results suggesting auditory cortex as a locus where the neural representation of vocalization sounds generalizes over production variability is consistent with a recent study showing that neurons in the auditory cortex of ferrets show robust responses to vowel identity tolerant to manipulations of various vowel features 38 .…”

BioCPPNet: Automatic Bioacoustic Source Separation with Deep Neural Networks

“…However, this approach to data preprocessing is not always feasible for non-human animals. For instance, while the fundamental frequency (F0) of macaque vocalizations is on the order of 0.1-1 kHz 16 , which suggests that they could be treated similarly as human speech in terms of downsampling to reduce computational costs, the mean maximum F0 of bottlenose dolphin (Tursiops truncatus) signature whistles can range from 9.3-27.3 kHz 17 , which means that with sampling rates of 96 kHz, even the lowest order overtones approach or exceed the native Nyquist frequency. This is further exacerbated in animals such as bats, which can emit broadband calls with dominant frequencies ranging from 11-212 kHz 18 .…”