With the world’s third largest exclusive economic zone (EEZ), Australia has strategically developed its blue economy over the past decade, comprising shipping (ship building and port expansion), offshore energy (oil, gas, and renewables), tourism, fisheries, and aquaculture (Australian Government, 2013). Western Australia, in particular, has a strong history of offshore energy development, ranging from oil and gas off the northwest coast to upcoming windfarms off the southwest coast ^{1 , 2} . Moreover, the growth of Western Australian ports is planned for both shipping and maritime defense ^{3 , 4} . Sustainability of the blue economy hinges on careful environmental planning to safeguard Western Australia’s marine biodiversity.

The Western Australian offshore environment provides habitat to ten species of baleen whale: common minke whale (Balaenoptera acutorostrata), Antarctic minke whale (B. bonaerensis), sei whale (B. borealis), Bryde’s whale (B. edeni), blue whale (B. musculus), Omura’s whale (B. omurai), fin whale (B. physalus), pygmy right whale (Caperea marginata), southern right whale (Eubalaena australis), and humpback whale (Megaptera novaeangliae). Of these, the southern right whale and the blue whale are listed as endangered in Australia ⁵ . The blue whale comprises two subspecies: the Antarctic blue whale (B.m. intermedia) and the pygmy blue whale (B.m. brevicauda). While the southern right whale and the Antarctic blue whale mainly occur seasonally along the southwestern and southern coasts of Western Australia, the pygmy blue whale occurs mainly along the entire Western Australian coast, annually migrating from the southern feeding grounds to the northern breeding grounds. The Blue Whale Conservation Management Plan (Australian Government, 2015) requires that anthropogenic threats be demonstrably minimized to aid the recovery of this species.

Effective conservation management requires adequate monitoring of abundance and distribution. Observation methods include visual surveying and passive acoustic monitoring (PAM) (e.g., Verfuss et al., 2018). Visual surveys may be done from shore for species traveling close to shore or require boats, planes, or drones to travel along predesigned transects or along the coast, with observers counting or photographing animals per time or area. Visual surveys can be expensive to carry out and can be challenging in bad weather or poor light. Additionally, some (sub-) species are difficult to visually tell apart in the field (e.g., Bryde’s and Omura’s whales, Antarctic and pygmy blue whales). Often in combination with visual surveys, PAM can be done by towing acoustic receivers along line transects. In some situations, recorder packages may be moored on the seafloor. PAM is independent of light and (mostly) weather conditions. A drawback of PAM is that it only detects animals when they vocalize. However, PAM can tell (sub-)species apart by their stereotypical sounds (in particular, songs).

Blue whales produce songs that are stereotypical to subspecies. The Antarctic blue whale song contains the Z-call (named after its Z-shape in spectrograms), consisting of a ∼9 s constant-wave sound at ∼28 Hz, followed by a ∼1 s downsweep from ∼28 Hz to 18 Hz, ending in a ∼8 s constant-wave sound at ∼18 Hz (e.g., Gavrilov et al., 2012; Miller et al., 2014). Five different populations of pygmy blue whales were distinguishable by their stereotypical songs in the Indian Ocean (Leroy et al., 2021). The Eastern Indian Ocean pygmy blue whale (EIOPBW) sings songs with up to three units: a ∼40 s constant-wave unit, followed by a ∼20 s upsweep, followed by another ∼20 s constant-wave unit, all with fundamental frequencies in the range 18–23 Hz (Gavrilov et al., 2011; Jolliffe et al., 2023). Evidence suggests that only males sing, likely as part of breeding behavior (McDonald et al., 2001). On the other hand, both males and females produce non-song sounds (e.g., Cusano et al., 2022).

Blue whale non-song sounds are typically short (up to a few seconds), low in frequency (<100 Hz) and can be constant-wave or frequency-modulated. Non-song sounds have been associated with various functional behaviors, including feeding and mating (McDonald et al., 2001; Oleson et al., 2007a; Oleson et al., 2007b; Lewis et al., 2018; Schall et al., 2020). The most reported non-song sound is a downsweep, often referred to as D-call. Given that all demographic cohorts produce non-song sounds, their detection is desirable for species monitoring. However, with the ever-increasing amount of PAM data collected, automated tools are needed to detect them.

Automated methods to detect and classify EIOPBW sounds exist for both song and D-calls. Gavrilov and McCauley (2013) optimized an EIOPBW song detector based on spectrogram correlation of the first and third harmonics of the song unit type II. Less than 5% of the true sounds were missed. This detector is part of the CHORUS software package (Gavrilov and Parsons, 2014) ⁶ . Guilment et al. (2018) implemented a trainable dictionary-based algorithm with sparse representations to detect and classify downsweeps amongst other mysticete calls, with great success despite the variability of any specific call. Torterotot et al. (2019)] built on Guilment et al. (2018) and improved their detector by post-processing the detections to specifically lower the false positives for Antarctic blue whale D-calls. They were able to reach an average of a single false positive per hour on their datasets. Miller et al. (2023) used a DenseNet architecture (Huang et al., 2017) to detect downsweeps from fin and blue whales off Antarctica. The neural network detection probability outperformed manual analysis by a human expert by more than 20% and 5% for low-medium and high signal-to-noise ratios, respectively. To detect EIOPBW D-calls with neural networks (similarly to Miller et al., 2023 or Rasmussen and Širović, 2021) or for dictionary-based approaches (Guilment et al., 2018), a training database of these specific sounds is needed, so the detector parameters may be tuned.

While the aim of our study had been to build and optimize a detector for EIOPBW non-song sounds, including downsweeps, detector performance indicated great variability of downsweeps, so instead, we employed clustering approaches, and by the spatio-temporal occurrence of downsweep types off Western Australia discuss the challenges and contextual nuances in assigning downsweeps to species.

The EIOPBW non-song sound detector (spectrogram correlator) outputs nearly 129,000 calls, which were manually sorted into classes. Of these, 120,685 were downsweeps, including 6,997 of type EIO1 and 407 of type EIO4. Detections of potential EIO2 and 3 sounds were much rarer and often confounded with ship noise or EIOPBW song units, and therefore discarded from further analysis. The following sections illustrate the variability of EIO1 sounds and the gradual morphing of downsweeps.

We identified morphing cases of different downsweeps starting from the EIOPBW non-song vocalization EIO1 reported in Recalde-Salas et al. (2014). When relevant, the different sounds in the morphing plots that resembled a published spectrogram of a species were pointed out.

The main goal of this work was to build and optimize a detector for EIOPBW non-song calls, which are mostly of downsweeping type, to further study this species’ geographic and seasonal pattern of occurrence. We tried two automated detectors: 1) a simple spectrogram correlator based on stencils of confirmed (by simultaneous visual and acoustic survey; Recalde-Salas et al., 2014) EIOPBW non-song sounds, and 2) a neural network that had previously been trained on general blue whale D-calls (Miller et al., 2023) followed by an automated clustering algorithm. Upon manual checking of the auto-detections and clusters (and confirmed by manual detections, feature measurements, and PCA), we found that downsweeps exhibit great variability, do not cluster well, and instead are graded (i.e., lie along an acoustic continuum). We provided several examples of such gradations in call duration; bandwidth; start, end, and maximum frequency; presence of overtones; frequency modulations; and “decorations”.

Many other authors have noted the great variability of blue whale downsweeps in other parts of the world (e.g., Rankin et al., 2005; Berchok et al., 2006; Oleson et al., 2007a; Oleson et al., 2007b; Torterotot et al., 2023), without grading them or lining them up into a continuum. As we compared our downsweep variants to spectrograms from the literature, we noticed that along the continuum, downsweeps morph through features that have been described for other (non-blue whale) species. This raises questions about the crudeness of describing sounds from spectrographic features alone and ultimately, about our ability to generalize assignation of (non-song) sounds to species in the absence of knowledge about the ecological and behavioral contexts at the time and location of acoustic recording.

The most common call type in the observations by Recalde-Salas et al. (2014) was EIO1, also frequently observed in our datasets. This call is graded in duration, frequency, presence of overtones, and decorations. For example, the EIO1 variant in Figure 3Ad is similar to the DS1 call spectrogram published for sei whales and the EIO1 variant in Figure 3Be is similar to the DS1H call also published for sei whales (Cerchio and Weir, 2022) (Figures 14A–D). The former call type was also recorded by Tremblay et al. (2019) from sei whales, and the latter call type by Rankin and Barlow (2007) and Cusano et al. (2023) from sei whales. The EIO1 variant of Figure 3Bb resembles the humpback whale “muah” (Recalde-Salas et al., 2020) and the humpback whale B-call (D'Souza et al., 2023); variants in Figure 3Ba Ca look similar to published humpback downsweeps (call type D, Fournet et al., 2015; call type G; Epp, 2019; call type B; Saloma et al., 2022; and one call in Indeck et al., 2020). The variant in Figure 3Bc resembles the downsweep recorded from Antarctic minke whales (Casey et al., 2022). The variants with extended lips resemble calls from humpback whales (“Low Hum”, Epp, 2019; “Descending Moan” and “Wup”, (Fournet et al., 2015); “Eaw” and “Modulated call”, Cusano et al., 2020; “Screech”, (Dunlop et al., 2007), or right whales (“downcall” and “hybrid”, Webster et al., 2016), or sei whales (“arch-call”, Cerchio and Weir, 2022). Figure 3Dc, Di compare with calls from baleen whales geographically as far away as bowhead whales in the Arctic (Thode et al., 2017).

How then can calls be assigned to species in the absence of visual validation or relevant ecological contextual knowledge? We explored 1) acoustic co-occurrence of downsweeps with species-stereotypical calls, 2) acoustic context provided by sounds before and after, and 3) geographical and seasonal occurrence of downsweeps. While downsweeps were sometimes recorded together with Antarctic minke whale bioducks (Dominello and Širović, 2016) or Omura’s whale calls (Browne et al., 2024), they occurred at variable time relative to each other and were, hence, not biphonations. With ten species of baleen whale occurring around Australia, most of which are known to migrate annually between colder (southern, in summer) and warmer (northern, in winter) grounds, co-occurrence has been noted frequently (e.g., Erbe et al., 2015), and so, sounds recorded at the same time do not have to come from the same species. Given almost all of our data were recorded with a single, omni-directional sensor, the two calling animals might not even have been at the same location. Our sonobuoy recordings provided bearing information to the calling individuals and EIO1 downsweeps followed the track of an EIOPBW singer. It is possible that the same individual produced both song and downsweeps, or that a singer and non-song producer traveled closely together in a small cohort. In the latter case, given the two sound types were tracked together for 80 min, it is likely that the cohort consisted of individuals of the same species and therefore, that those downsweeps were indeed made by an EIOPBW.

Acoustic context can be derived from other sounds occurring before or after a downsweep. Some of our downsweep detections were part of patterned sequences or songs. Humpback whales produce complex songs, using a great variety of units in long and hierarchical patterns (Payne and McVay, 1971) and some of our downsweeps were part of these complex patterns, but not all. For downsweeps in simpler or no patterns, identification to species was thus inconclusive.

Ecological and behavioral context in a given region can also inform species identification. For instance, seasonally occurring downsweeps during known migratory periods of PBW may be attributable to the species with a high level of certainty in locations where other species that produce downsweeps are very unlikely to occur. The geographic and seasonal occurrence of downsweeps resembling EIO1 from Recalde-Salas et al. (2014) reported here is an example where knowledge of the migratory timing of PBWs corroborates downsweep attribution to the species. For downsweeps not resembling EIO1, looking at the geographic and seasonal occurrence, however, was inconclusive as they could not be separated into clusters.

More work needs to be done before downsweeps as single spectrographic features, in the absence of regional ecological and behavioral context, can perhaps be used routinely in environmental mitigation and monitoring plans. First, beginning with automated detectors, there will always be a trade-off between precision and recall, between false alarms and missed detections. More specifically, spectrogram correlation detectors are known to suffer from a lack of variability in the detected signals due to fixed templates (Socheleau et al., 2015). Their advantages are that they are intuitive, simple, and quick to set up; they do not require the creation of a training database for dictionary-based methods or neural networks. Second, sorting detections into classes is prone to biases. The UMAP algorithm was unable to cluster downsweeps—partly because downsweeps are graded. Moreover, similar call types were found in different clusters because of different ambient noise in the recordings. Even for high signal-to-noise ratio examples, UMAP may cluster based on noise features as the algorithm is influenced by both local and global structures in the data (McInnes et al., 2018). Statistical clustering techniques could have been applied to quantify the lack of clusters in the data. However, these methods also suffer from drawbacks such as sensitivity to noise, dependence on predefined parameters, and the potential to detect spurious clusters that do not reflect meaningful differences in the data. Furthermore, clustering algorithms may impose artificial groupings based on mathematical assumptions rather than perceptual or biologically relevant differences in the sounds. Therefore, given that neither PCA or UMAP revealed clusters, a visual inspection was considered sufficient to conclude that the calls do not naturally cluster. Finally, manual sorting of detections, apart from being overwhelming, is biased by human perception (Leroy et al., 2018; Nguyen Hong Duc et al., 2021; Dubus et al., 2024). One solution to averaging out the bias in the human error could be to have citizen scientists annotate the data (Nguyen Hong Duc et al., 2021; Dubus et al., 2024). This solution has limitations, however, because the results may be prone to greater error and lower precision, and the bias will still be present due to, for example, how the citizen scientists were trained, their previous experience in annotating, and the material they use for training. Additional improvements, in particular for highly variable calls like downsweeps, might be achieved by soft instead of hard labels for sounds, such as probabilities for belonging to a type, population, or species, rather than discrete call categories.

Given the similarities of call types across different species, we might need to improve our ways of measuring and describing these sounds. Transforms other than the Fourier transform (e.g., wavelet transforms; Urazghildiiev and Clark, 2007; Mouy et al., 2008) might capture different features and separate calls differently. Correcting for the effects of the sound propagation environment could help (e.g., removing echoes, frequency-dependent absorption, and dispersion; Erbe et al., 2022), but the location of the calling animal is typically unknown in recordings from single, autonomous deployments. Furthermore, a common library of audio sounds (Miller et al., 2021; Parsons et al., 2024) with associated metadata (site, time of the year, sampling frequency, supposed category) could enable researchers to compare their annotations and increase their training, testing, and reference database.

Finally, tools derived from phonetic science might improve researchers’ accuracy in identifying species producing specific call types. Phonetic science focuses on a combination of physical properties, perception, and transmission of sounds and language (Kortmann, 2020), in the context of recognizing, categorizing, and understanding sounds. Species could have differences in pitch and loudness that could contribute to manual downsweep identification; albeit for sounds having frequencies below those humans are sensitive to, these would need to be sped up during playback. Although perception of sounds is inherently biased, theories and tools on speech perception and sound production in humans may help manual categorization of sounds accounting for how these are processed by the brain. Until further advances can be made to allow full automation of downsweep detection as single spectrographic features, the extent of interpretation of results and their attribution to species will need to be informed by regional ecological and behavioral knowledge and manual reviewing of acoustic data by experienced personnel in the identification of sounds spectrographically and phonetically. Then, an integrated approach drawing from different disciplines could be a critical pathway to progress automated classification of downsweeps at the species level.

In conclusion, we showed that downsweeps do not cluster; instead, they are graded and morph along a continuum of acoustic features. We also demonstrated the challenges in classifying downsweeps to species, with similar downsweep types having been reported from diverse species of mysticete whales. Hence, not all low-frequency downsweeps (<100 Hz) are blue whales. Even humpback whales have downsweeps to below 50 Hz in their songs. While automated passive acoustic monitoring for environmental management would ideally detect non-song sounds in addition to species-stereotypical song (in order to not only increase the probability of detection but also monitor the non-singing demographics), reliance on downsweeps alone is marred with several challenges—at this stage.