We analyzed 682 specimens from 20 museum collections, including 192 females and 490 males (File S1). We personally examined the holotypes of C. c. immaculata, C. c. iterata and C. c. sclateri; the syntypes of C. c. immaculata; the lectotypes of C. c. cinerascens and C. napensis; and the paralectotypes of C. c. cinerascens and C. c. sclateri. We visually assessed plumage color variation across several features, including auriculars, back, belly, breast, chin, coverts, crissum, crown, feathers, flanks, head, interscapular patch, bill, lores, neck, shoulders, supercilium, tail, tail tip, thighs, throat, and uropygium, using a soil chart color catalogue (Munsell 1994). All measurements were taken on the left side of the specimens by VC for consistency.

We measured nine features using scales (0.1 mm) and electronic calipers (0.01 mm) (Eck et al. 2011). These features included: culmen length (BNdist), from the distal edge of the nostril to the tip of the bill; wing chord (Wchord), from the carpal joint to the tip of the wing; tail length (T1), from between the innermost rectrices to the tip of the longest feather; tarsus length (Tar1), from the joint between the tarsus and toes to the intertarsal joint; tail graduation (Tgrad), from the tip of the longest to the shortest feather; and tail tip length (R1), corresponding to the whitish tip of the outer rectrices, when present.

We analyzed 347 recordings (File S2) and classified vocalizations into notes (continuous lines separated by silence on a spectrogram, shown as near-zero intensity in oscillograms) and phrases, which consist of one or more notes (Carneiro et al. 2019). Phrases, which form loudsongs, are separated by longer intervals than those between notes. Recordings were obtained from the Fonoteca Neotropical Jacques Vielliard (FNJV), Macaulay Library of Natural Sounds (Cornell Laboratory of Ornithology, Ithaca), Wikiaves (www.wikiaves.com.br), and Xeno-canto (www.xeno-canto.org). All .mp3 files were converted to the .wav format before analysis. We examined spectrograms of vocalizations using Raven Pro 1.6.5 (K. Lisa Yang Center for Conservation Bioacoustics at the Cornell Lab of Ornithology 2024).

The analyzed robust vocal characteristics (Zollinger et al. 2012) included bandwidth at 90% (frequency range that contains the central 90% of the acoustic signal’s energy), center frequency (frequency of the pixel with the median power within a selection), center time (the point in time within a selection at which 50% of the sound energy has occurred earlier), frequency contour percentile at 5% (frequency 5% measurements for spectrogram slices), frequency contour percentile at 95% (frequency 95% measurements for spectrogram slices), maximum and minimum peak frequency contours, and time at 5% (proportion of the selection at which 5% of the sound energy has an earlier time) and at 95% (proportion of the selection at which 95% of the sound energy has an earlier time). We also examined note duration (s) and number (which included raspy, clear and raspy-clear notes), pace (note/s), song duration (s), the number of consecutive raspy notes, and the proportion of raspy, clear and raspy + raspy-clear notes. Additionally, we considered one categorical feature: note shape. Vocalizations in each recording were analyzed, up to a maximum of 10 loudsongs. Overlapping notes or songs were excluded. We used the default parameters in Raven Pro 1.6.5, except for the frequency grid spacing, which was set to 10.8 Hz. This value refines the frequency range (172 Hz in default mode) by accounting for the uncertainty principle in analyzing vocalizations (Vielliard and Silva 2010). This principle states that increasing the precision of one parameter reduces the precision of a related parameter, such as temporal versus frequency scales in a spectrogram. To obtain more accurate frequency measurements, we accept lower time precision, and vice versa. All spectrograms were analyzed at the same scale (0.1 s/cm and 0.4 kHz/cm). Brightness and contrast were adjusted for each spectrogram without losing information. Frequencies were measured using frequency sweeps (selection spectrum), while time was measured with oscillograms (waveforms). Spectrograms were generated using warbleR (Araya-Salas and Smith-Vidaurre 2017) and Raven Pro 1.6.5.

Based on our previous understanding of vocal variation within C. cinerascens populations, we identified five potentially separate lineages for all subsequent analyses: the northern C. c. cinerascens (including C. c. immaculata) found north of the Amazon River; and four southern populations: C. c. sclateri (east of the Andes extending to the Ucayali River), the Ucayali-Madeira interfluve population, the Madeira-Tapajos interfluve population, and C. c. iterata (east of the Tapajos River).

We tested for premises at α = 0.01 before performing a Multivariate Analysis of Variance (MANOVA) to analyze the means of morphometric and vocalization measurements across multiple dependent variables simultaneously. Analysis of Variance (ANOVA) was used to compare measurements between groups. If these differences were statistically significant, we applied a Tukey post-hoc test with an adjusted p value to determine significant differences within taxa. To analyze and classify the data, we employed Linear Discriminant Analysis (LDA). The dataset was preprocessed to ensure it met the assumptions of LDA, including checking for multivariate normality and homogeneity of covariance matrices among classes. Due to collinearity, bandwidth 90% was removed from the analyses. For better interpretability, graphical representations were created overlaying 95% confidence ellipses for each class to illustrate the separation between them. To estimate whether the ranges did not overlap with larger sample sizes we assumed: X_a+t_a SD_a≤X_b-t_b SD_b in which X are means and SD are the standard deviations of the populations with the smallest (a) and the largest (b) set of measurements; t_i is the t score at the 97.5 percentile of the t distribution for n – 1 degrees of freedom (Isler et al. 1998). All analyses were conducted within the R environment (R Core Team 2023).

All recordings were standardized to a sampling rate of 44.1 kHz and 16-bit depth in .wav format using Audacity 3.1 (Audacity Team 2021) before analysis. We visually inspected vocalizations in Raven Pro and measured the acoustic parameters by analyzing spectrograms generated with a 1024-point FFT, Hann window, and 90% overlap.

To supplement manual measurements and capture subtle multidimensional acoustic variation, we extracted 1024-dimensional feature embeddings from audio recordings using the BirdNET algorithm, a deep neural network (DNN) based on a 157-layer Residual Network (ResNet) architecture with 27 million parameters trained on millions of globally distributed bird vocalizations (Kahl et al. 2021; McGinn et al. 2023).

Audio clips were created from recordings that were manually annotated for start and end times, as well as the minimum and maximum frequencies, previously identified for each loudsong using Raven Pro selections. These segments were automatically extracted with the librosa library to produce standardized 3-second clips centered on the vocalization. A Butterworth filter, implemented in SciPy, was used to restrict frequencies to the annotated range and to reduce external noise. The clips were exported as 48 kHz .wav files, ensuring consistency in time and spectrum for BirdNET processing.

When an audio segment is processed by BirdNET, intermediate layers of the network encode acoustic structure into a numeric vector – known as a feature embedding. These are extracted using the public embeddings.py script from the BirdNET-Analyzer repository (model version 2.1). This process produces numerical vectors from the penultimate layer of the DNN, which encode high-dimensional acoustic information from each vocalization into a standardized format. BirdNET then processes the loudsongs, converts them into spectrograms, and transforms them into 1024-element vectors. They summarize complex temporal and spectral information from the spectrogram, including frequency modulation, note shape, and temporal rhythm, in a way that is optimized for capturing biologically meaningful variation (Ghani et al. 2023).

We used Uniform Manifold Approximation and Projection (UMAP) (McInnes et al. 2018) to reduce the dimensionality of the embeddings to two dimensions for visualization. Labels were assigned based on filenames, grouping vocalizations into the five predetermined populations. The resulting structured projection preserved similarities between populations and was visualized as 2D scatter plots with color-coded groups. To assess whether populations could be statistically distinguished, we trained a Support Vector Machine (SVM) (Cortes and Vapnik 1995) and compared its performance to a random classifier using the “most frequent” strategy (Sanchís-Pedregosa et al. 2011), which predicts all vocalizations as the same class. Model performance was evaluated through repeated random subsampling validation (100 iterations), using a 60/40 train-test split (Lakdari et al. 2024). The assessment included overall accuracy and class-specific precision, recall, and F1-score metrics (File S3). This combined approach enabled us to compare the natural data structure revealed by unsupervised analyses with the class separability under a supervised scenario, providing a comprehensive view of the utility of vocalization embeddings for distinguishing different lineages within the C. cinerascens species complex.

We employed a multi-criteria diagnosability framework, weighting qualitative characters in both plumage and loudsong structure equally with quantitative measurements. Following McKitrick and Zink (1988), we applied the conventional threshold that 95-99% of the individuals must be assignable using either defining traits. For the C. cinerascens complex, diagnoses were based on qualitative plumage differences (e.g., light and dark gray morphotypes). Loudsongs were diagnosed using two criteria: continuous characters required non-overlapping value ranges, while categorical characters had to be unambiguously identifiable aurally or visually in spectrograms and show no clinal variation (Isler et al. 1998). For C. c. iterata we used a unique combination of the above. Thus, each proposed taxon was defined by consistent, quali-quantitative character states that meet the diagnosability criterion.

We diagnosed two morphotypes based on the overall plumage color and patterns of both females and males. These morphotypes are separated by the Pastaza-Marañon-Solimões-Amazon Rivers system. The dark brownish (olive 5Y 4/4) female morphotype occurs south of these rivers. It is characterized by a white interscapular patch, buffy fimbriae on the external wing coverts, and wide (> 3.5 mm) buff-tipped tail feathers. In contrast, the lighter olive-brown (olive yellow 2.5Y 6/8) morphotype is found north of these rivers and lacks the white and buffy markings. The dark gray (very dark gray N/3) male morphotype is distributed south of the river system and features a white interscapular patch, white fimbriae on the external wing coverts, and whitish wide tail tips. North of this system, the light gray (gray N/5) male morphotype also differs from the previous one by lacking these white markings (Fig. 1).

Significant differences. All measurements showed significant variation among females of the five populations (MANOVA, F_1,4 = 6.2, p < 0.001), except for culmen length and tail graduation (Table SS1). Some traits also varied significantly within populations, particularly when comparing northern and southern morphotypes (Table SS2). Males from the five populations also differed in measurements (MANOVA, F_1,4 = 12.3, p < 0.001), with all differences being significant (Table SS3). These within-population differences also reflected the northern and southern morphotypes (Table SS4).

Diagnosis. There was an overlap in measurements for both females and males across all five populations (Tables 1, 2; Figs S1, S2). Therefore, we found no diagnostic morphometric features to reliably identify them.

Multivariate space. For females, the LDA model achieved an overall accuracy of 64.8%. The 95% confidence interval for this accuracy ranges from 55.9% to 73.0%. The Kappa statistic, which accounts for agreement by chance, was 0.523. For males, these values were similar: the overall accuracy was 65.0%, with confidence intervals from 59.7% to 70.1%, and Kappa was 0.469. Overall, the LDA models demonstrated moderate accuracy and Kappa values, indicating reasonable class separation and prediction ability (Fig. 2A, B). The linear combination of predictor variables used to develop the LDA model’s decision rules is listed in Table SS5.

There was limited information about the sex of the birds in the recordings. Details regarding the absence of playback were scarce (found in only 12% of the recordings), and only one recording (XC 138894) from Viruá National Park in Caracaraí, Roraima, Brazil, explicitly indicated the use of the technique. Therefore, all vocalizations were included in our analyses. The loudsong of this species complex features disyllabic phrases with one or two notes, depending on the population, and consists of a variable number of phrase repetitions. Qualitatively, the isolated notes are raspy or clear, but in one population (Andes-Ucayali interfluve), these notes merge into a single raspy-clear note.

Significant differences. The measurements across all five populations varied significantly (MANOVA, F_1,4 = 24.4, p < 0.001), but differences in bandwidth 90% were not statistically significant between populations (Table S6). Some 94 measurements also differed significantly within populations (Table S7), but with considerable overlapping ranges.

Diagnoses. Six vocal measurements overlapped among the five populations, but for nine other traits, non-overlapping was consistent from 5-8 vocal features, except for cinerascens-iterata, and cinerascens-Ucayali-Madeira and Ucayali-Madeira-iterata pairwise comparisons (Fig. S4; Table 4). Within the Ucayali-Madeira population, loudsongs selected from four recordings lacked the diagnostic two-note initial raspy sequence. However, in three of these (ML135178, XC142165, and WA1010195), loudsongs containing this diagnostic introductory sequence were also present in the remaining loudsongs. Only in the recording ML88029, from the Pilon Lajas Biosphere Reserve in Bolivia, were the two available loudsongs entirely lacking this sequence. In these recordings, eight loudsongs did not have the initial consecutive raspy notes, representing 3% of the examined vocalizations (n = 265), which remains within our ≥95% diagnosability threshold.

Multivariate space. The LDA model achieved an overall accuracy of 86.7%, with a 95% confidence interval from 79.3% to 86.6%. The Kappa statistic, which adjusts for chance agreement, was 0.825. Overall, the LDA model showed strong performance with high accuracy and Kappa values, indicating effective class separation and reliable predictions (Fig. 3). The coefficients reveal that the proportions of types of notes and note duration have the highest weight for the first discriminant function, suggesting they play a major role in distinguishing the groups (Table S8). During the visual inspection of the spectrograms, however, we could only identify four patterns among the five C. cinerascens populations (Figs 4, S5):

Cercomacra cinerascens cinerascens (including C. c. immaculata and C. c. iterata). Mostly (96.9%) consists of two notes: a clear note and a raspy note. The remaining phrases are made up of a single note where the clear and raspy elements are combined but still distinguishable. The first note of this song can be clear (62.4%), raspy (11.6%), or may lack a time interval between the raspy and clear notes (or raspy/clear element; 26%), effectively making it a single note. In this case, this occurs mainly (99.2%) at the start of the song and rarely elsewhere. Clear notes (52.5%) are more common than raspy notes (47.5%). The number of phrase repetitions per loud song ranges from 3 to 29, with an average of 11.2 ± 4.9. The alternating pattern of clear and raspy notes visually characterizes it.

Cercomacra cinerascens sclateri. Diagnosed by phrases composed exclusively of raspy-clear notes. The number of phrase repetitions per song ranged from 3 to 18, with an average of 6.9 ± 4.1. The note duration and pace of this loudsong do not overlap with others. It can be visually distinguished by a single note composed of clear and raspy elements.

Ucayali-Madeira Interfluve. Mostly (68.8%) consisted of phrases with two raspy notes and one clear note. When a phrase had both raspy and clear elements, it appeared at the beginning of the song (31.2%), and rarely (3%) in another position. The loudsong typically started with two raspy notes (or a raspy-clear element followed by two raspy notes), followed by a clear note. There is a higher occurrence of clear notes (55.9%) compared to raspy notes (44.1%). The number of phrase repetitions per loudsong ranged from 2 to 20, with an average of 7.4 ± 2.6. It can be visually identified by the two or three raspy notes before the appearance of the first clear note.

Madeira-Tapajos Interfluve. Diagnosed by the presence of raspy notes only, since the homologous clear note of the other loudsongs is represented by a raspy element. Most of the phrases (85.7%) were produced in even numbers. The number of phrase repetitions per song ranges from 2 to 10, with an average of 6.2 ± 1.6. It can be visually identified by the raspiness of the homologous clear note present in the other species’ loudsongs.

The geographical distribution of loudsong C. c. cinerascens is north of the Pastaza-Marañon-Solimões-Amazon system, but also south of the Amazon River east of the Tapajos River, applying to C. c. cinerascens, C. c. immaculata and C. c. iterata. Loudsong C. c. sclateri is limited south of the Pastaza River in Ecuador and Peru, bordering the Andes to the west and the left bank of the Apurimac and Ucayali Rivers to the east. Loudsong Ucayali-Madeira is restricted to the eastern banks of the Tambo and Ucayali Rivers, extending east to the left bank of the Mamoré River, exclusively given by this unnamed population. Loudsong Madeira-Tapajos is found within the Madeira-Tapajos Rivers and only recorded within this interfluve (Fig. S5).

These vocal patterns were confirmed using BirdNET-derived embeddings. The resulting projection showed significant overlap among all five populations, with some evidence of distinct clustering, especially for C. c. sclateri (Fig. 5). The SVM model achieved a high accuracy of 0.90, with population-specific accuracies ranging from 80 to 100%. Misclassifications occurred frequently, particularly for the loudsongs from the eastern banks of the Tapajos River, reflecting their overlapping distributions in the embedding space. A random classifier performed considerably worse, underscoring the embeddings’ limited but non-random discriminatory power (Table S9). The distinction of loudsongs among populations was not very pronounced and the model had a moderate balanced ratio between precision and recall, ranging from 0.79 to 1.00. These results suggest that vocal divergence among populations is high for loudsong C. c. sclateri, which achieved perfect precision and recall.

Based on the combination of vocal and morphological analyses, we identified consistent quantitative and qualitative diagnostic characters (Tables 4, 5). In this context, the diagnosis is unambiguous, meaning that these features – either alone or combined – reliably distinguish the five proposed taxa. These species are separated by major Amazonian rivers, suggesting allopatric or parapatric distributions with little or no contact zones (Fig. 6). Although secondary contact might theoretically occur in river headwaters, we found no evidence of intergradation or hybridization in these regions, supporting the interpretation that these populations are distinct and geographically structured. As a result, we recognized five species, two of which lack previously available names.