Paper deep dive
Structure Across Voices: Comparing acoustic-event type accumulation and sequence dependence across four vocal repertoires using frozen audio encoders
Mudit Sinha, Sanika Chavan
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Vocal repertoires can differ in acoustic-event type accumulation and temporal organization, yet direct comparison is difficult because corpora use different native events and unequal amounts of sequence. We compare sperm whale codas, human speech phones, Bengalese finch syllables, and common marmoset calls using the same frozen-audio-encoder procedure while matching event count and local sequence opportunity. Whale shows the fastest type accumulation; Finch shows the strongest immediate dependence and repeated-subsequence recurrence. Physically interpretable acoustics recover complementary parts of this profile, continuous analyses without clustering support broad Whale acoustic coverage, and source- and position-preserving nulls retain both Finch order effects. Extending predictive context shifts the comparison toward Whale. Thus repertoire differences depend on the acoustic property and temporal scale measured rather than forming a single hierarchy.
Tags
Links
- Source: https://arxiv.org/abs/2609.16612v1
- Canonical: https://arxiv.org/abs/2609.16612v1
Trouble viewing inline? Open PDF directly â
Full Text
45,855 characters extracted from source content.
Expand or collapse full text
Structure Across Voices: Comparing acoustic-event type accumulation and sequence dependence across four vocal repertoires using frozen audio encoders Mudit Sinha a and Sanika Chavan b Independent Researcher, San Bruno, California 94066, United States (Dated: 5 September 2026) Vocal repertoires can differ in acoustic-event type accumulation and temporal organization, yet direct comparison is difficult because corpora use different na- tive events and unequal amounts of sequence. We compare sperm whale codas, human speech phones, Bengalese finch syllables, and common marmoset calls using the same frozen-audio-encoder procedure while matching event count and local sequence opportunity. Whale shows the fastest type accumulation; Finch shows the strongest immediate dependence and repeated-subsequence recur- rence. Physically interpretable acoustics recover complementary parts of this profile, continuous analyses without clustering support broad Whale acoustic coverage, and source- and position-preserving nulls retain both Finch order ef- fects. Extending predictive context shifts the comparison toward Whale. Thus repertoire differences depend on the acoustic property and temporal scale mea- sured rather than forming a single hierarchy. Š2026 Acoustical Society of America. [https://doi.org(DOI number)] [XYZ]Pages: 1â12 I. INTRODUCTION Vocal repertoires contain recurring acous- tic events, but the event itself differs across corpora: sperm-whale codas, human speech phones, Bengalese-finch syllables, and common- marmoset calls. These units differ in dura- tion and internal complexity; imposing equal- duration segments would introduce a different assumption about where meaningful events be- gin and end. We therefore preserve each cor- pusâs annotated event scale and ask what re- mains comparable after matching the opportu- nity to observe diversity and order. This match- ing is necessary because the corpora also differ greatly in size and contiguous sequence length. We ask two linked questions: how quickly do new acoustic-event groups continue to ap- pear, and how strongly does temporal context constrain the event stream? Temporal organi- zation is examined through immediate depen- dence, repeated-subsequence recurrence, and the a muditsinha01@gmail.com b sanikac10@gmail.com additional predictive value of the event two po- sitions back. We do not assume the native units are biologically homologous; we compare them using a common procedure for describing recorded acoustic events and matched observa- tion opportunity. Whale is limiting at 1,501 codas. We re- tain all Whale events and sample each 12,000- event reference pool into the same 1,501-event, 113-block profile, equalizing both event count and opportunities for adjacent pairs and tri- grams. Frozen encoders then describe every event under the same signal-processing proce- dure, while groups are formed separately within each repertoire. Physical-acoustic and contin- uous analyses test whether the resulting pro- file exists outside the learned representation; source- and position-aware controls test alterna- tive explanations; longer-context prediction and a Whale click-timing perturbation test temporal scale and a concrete physical timing contribu- tion. J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 1 arXiv:2609.16612v1 [cs.SD] 15 Sep 2026 I. RELATED WORK Comparative vocal-sequence studies use en- tropy, transition statistics, and recurrence to quantify event dependence in dolphin whistles, humpback whale song, birdsong, and speech (Kershenbaum et al., 2016; McCowan et al., 1999; Morita et al., 2021; Sainburg et al., 2019; Suzuki et al., 2006). Such measurements depend on both event definition and available sequence: changing the observation unit changes the types and transitions that can be estimated. Type accumulation asks a complementary questionâ how rapidly observed categories continue to expandâfor which we use the HeapsâHerdan formulation (Heaps, 1978), treating âtypesâ as operational acoustic groups rather than biologi- cal categories. The four corpora package acoustic struc- ture at different native scales, and duration alone does not identify a uniquely correct cross- species unit. We therefore retain source-defined events and equalize observation opportunity rather than impose common-duration segmenta- tion (Kershenbaum et al., 2016; Sainburg et al., 2019). Sperm-whale codas additionally pro- vide a physically interpretable timing variable: their inter-click intervals vary across social units, clans, and contexts, and can be altered while retaining the recorded clicks (Gero et al., 2016; Sharma et al., 2024; Watkins and Schevill, 1977). Learned acoustic representations can reveal structure in animal vocalizations (Morfi et al., 2021; Sainburg et al., 2020), but here frozen en- coders are only a common description procedure; clustering remains repertoire-specific. We there- fore pair them with direct temporal and spec- tral measurements, continuous acoustic analy- ses without clustering, source-aware controls, and the Whale timing intervention. The com- parative gap is whether heterogeneous native event streams remain distinguishable when ob- servation opportunity and analysis procedure are matched. I. METHODS The comparison preserves each corpusâs na- tive acoustic event while equalizing the factors that would otherwise change the opportunity to observe acoustic diversity or event order: signal- processing procedure, clustering resolution, total event count, and block-length profile (Fig. 1). A. Corpora and source-defined acoustic events We retain each corpusâs native annota- tion rather than impose a common segmenta- tion. Whale contains 1,501 sperm-whale co- das, bounded click sequences conventionally de- scribed by ICI pattern (Gero et al., 2016; Sharma et al., 2024); hand-assigned coda type and tim- ing metadata are used only for interpretation. Reference pools contain 12,000 Buckeye conver- sational phones, 12,000 annotated Bengalese- finch syllables, and 12,000 MarmAudio calls (Koumura and Okanoya, 2016; Lamothe et al., 2025; Pitt et al., 2007). Native order is retained within recording tracks (Human), birds/songs (Finch), and recording days (Marmoset; caller identity unavailable). Table I separates available pools from the 1,501 events used per matched draw. TABLE I. Source corpora and matched sampling. Each reference pool contributes 1,501 events per draw in the Whale 113-block profile. Repertoire (native event) Available pool Per matched draw Source block Sperm whale (coda) 1,5011,501 (all)DateĂ Unit Human (phone) 12,0001,501recording track Bengalese finch (syllable) 12,0001,501individual bird Common marmoset (call) 12,0001,501recording day B. Common acoustic representation and within-repertoire clus- tering Eleven frozen configurations from VampNet, HuBERT, AVES-family, wav2vec 2.0, and Open- BEATs span speech-, general-audio-, music-, and bioacoustic-trained models (Baevski et al., 2020; Bharadwaj et al., 2025; Flores Garcia et al., 2023; Hagiwara, 2023; Hsu et al., 2021). Within a configuration, front-end treatment, temporal pooling, and encoder settings are identical across corpora, yielding one vector per event. K-means groups these vectors separately within each repertoire; K is the number of groups allowed, and K â 16, 24, 32 tests grouping resolution. Cluster identities are never aligned across corpora. Spherical and pooled- 2 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices Fair acoustic comparison across four vocal repertoires Source-defined events Sperm whale coda 1,501 available Human phone 12,000 available Bengalese finch syllable 12,000 available Marmoset call 12,000 available Common acoustic description same frozen encoder and temporal pooling for every repertoire one vector per event Within-repertoire K-means K = number of acoustic groups K = 16, 24, 32 groups fitted separately in each repertoire Matched rarefaction Whale: all 1,501 events References: 50 draws from each 12,000-event pool Every draw: 1,501 events | 113 blocks 1,388 adjacent pairs 1,275 trigrams Acoustic questions Type accumulation: how fast do new groups appear? Adjacent order: how predictive is the current event? Repeated sequences: how much recurrence? Longer context: does two-event history help? Matching event count and block lengths gives every repertoire the same amount of local sequence evidence. Native event definitions and recorded acoustic/acquisition differences remain corpus-specific and are analyzed separately. FIG. 1. Matched comparison. Native events are described with the same frozen-encoder procedure and clustered independently within repertoire. Every matched draw contains 1,501 events in 113 blocks, giving 1,388 adjacent pairs and 1,275 trigrams before breadth and sequence measurements are evaluated. center alternatives test clustering sensitivity (Supplementary Material). The groups are op- erational: in Whale they agree only weakly with historical CodaType and covary modestly with click count, duration, and normalized ICI, moti- vating the direct acoustic validations below. C. Matching event count and sequence opportunity Matching event count alone is insufficient because longer uninterrupted streams mechan- ically provide more transitions, trigrams, and repeated-subsequence opportunities. Whaleâs 1,501 codas are divided into 113 contiguous blocks of at most 16 events. Each of 50 refer- ence draws realizes the same block-length vector with non-overlapping spans inside native source groups; blocks never cross sources or duplicate events. Every comparison therefore contains 1,501 events, 113 blocks, 1,388 adjacent pairs, and 1,275 trigrams. Duration, spectrum, source composition, propagation, and recording hard- ware remain corpus properties and are addressed separately. D. Primary measurements With observation opportunity matched, we quantify two aspects of the acoustic event streams: how quickly distinct acoustic groups continue to accumulate and how strongly tem- poral order constrains those groups. Type accu- mulation measures the rate at which occupancy spreads across the fixed acoustic grouping as more events are observed. Immediate sequence dependence measures local, one-event predictive constraint. Subsequence recurrence measures re- peated multi-event ordering. Table I summa- rizes the scientific meaning of each quantity be- fore the estimator details below. a. Type accumulation. Let V (n) be the num- ber of distinct acoustic clusters observed after n events. The HeapsâHerdan exponent β summa- rizes V (n) â n β under the same accumulation J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 3 TABLE I. Primary measurements and their acous- tic interpretation. Type accumulation quantifies how quickly new acoustic groups continue to ap- pear. The two sequence measurements ask whether observed event order carries more local predictabil- ity or repeated subsequence structure than a within- block shuffle of the same recorded events and acoustic-group frequencies. Quantity EstimatorAcoustic question Type accu- mulation HeapsâHerdan β How quickly do previously unseen acoustic groups continue to appear as more events are observed? Immediate sequence depen- dence Observed- minus- shuffled first-order gain How much does the current acoustic group improve prediction of the next one beyond randomized order with the same recordings and group frequencies? Subsequence recurrence Observed/shuffled compression ratios Does the observed order reuse multi-event subsequences more than randomized order containing the same acoustic groups? convention, event budget, and K in every corpus (Heaps, 1978). Larger β means that new clusters continue to appear more rapidly as events accu- mulate. b. Immediate sequence dependence. We ask how much the current cluster assignment im- proves prediction of the next one beyond the same assignments in randomized order. Raw first-order context gain is the reduction in un- certainty about the next assignment after condi- tioning on the current assignment; the entropy terms use MillerâMadow finite-sample correction (Miller, 1955). We subtract the corresponding gain after shuffling assignments within blocks. Larger positive values therefore indicate more one-step predictive information in the observed order than in a sequence with the same assign- ment frequencies and block boundaries. c. Subsequence recurrence. We ask whether the observed event sequence contains more re- curring subsequences than the same assignments in shuffled order. RePair, LempelâZiv (LZ), and DEFLATE provide three observed/shuffled com- pression ratios (Deutsch, 1996; Larsson and Mof- fat, 2000; Ziv and Lempel, 1977). Ratios below one indicate that repeated subsequences make the observed sequence more compressible than its shuffle. The three compression readouts are treated together for the cross-corpus rank sum- mary. d. Cross-measurement profile. Because the three measurements have different numerical scales, their main cross-corpus comparison is the three-dimensional rank profile: each repertoire is ranked separately for type accumulation, shuffle- adjusted immediate dependence, and the mean of the three compression ranks. We do not use an aggregate rank to define an overall winner in the main text. An equal-weight mean-rank sum- mary is retained only as a secondary descriptive quantity in the Supplementary Material for sen- sitivity checks across analysis variants. E. Within-block shuffle baseline and uncertainty For order-sensitive measurements, cluster as- signments are shuffled within each matched block. This preserves the recorded events, clus- ter counts, block boundaries, and marginal fre- quencies while removing observed order, so static source and recording properties are shared by observed and null sequences. Position-varying acquisition or behavior can remain and is tested separately; type accumulation is unchanged by shuffling. Analytical stability is assessed across encoder choice, K, reference draw, coarsening, clustering geometry, RMS normalization, and alternative cluster constructions. Source uncertainty is kept separate by resampling the highest available na- tive source block: Whale Date Ă Unit, Human track, Finch bird, and Marmoset recording day (Supplementary Material). F. Physical acoustic validation beyond the learned representa- tion a. Physical-acoustic family decomposition. To ask which interpretable signal dimensions re- produce the encoder-derived profile, we analyze four acoustic families separately on the same 1,501-event, 113-block matched shell. Tempo- ral acoustics include duration, temporal centroid and spread, and envelope modulation. Spectral location and extent include peak frequency, cen- troid, bandwidth, and rolloff. Spectral shape includes spectral slope, flatness, and entropy. Acquisition-sensitive level measurements include RMS, peak level, clipping, and an SNR proxy. Spectral descriptors use the shared physical sup- port from 0 to 8 kHz, the maximum band avail- 4 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices able in every corpus. Each family is grouped and scored separately with the same three pri- mary measurements. Spearman Ď across the 12 repertoire-by-measurement mean-rank cells is used descriptively to ask whether a physical fam- ily preserves the whole encoder-derived profile rather than only one leading repertoire. Corpus metadata is retained as a non-acoustic compar- ison; no naive p-value is assigned to the 12 de- pendent rank cells. b. Continuous acoustic analyses without clustering. A complementary analysis stays in continuous physical-descriptor space. For acoustic coverage, 250 held-out query events are compared with ref- erence sets of 25, 50, 100, 200, 400, 800, and 1,200 events. Descriptor dimensions are stan- dardized on a pooled reference in which species contribute equally; nearest-neighbor distance is divided by the square root of feature dimension, and area under the distance-versus-log-sample- size curve summarizes how sparsely the acous- tic space is covered. A second version removes native-source means before the same calculation. These two standardizations are interpreted sep- arately rather than by comparing their abso- lute AUC scales. For local continuous predic- tion, a leave-one-native-source-out multi-output ridge model predicts the next descriptor vector from the current vector. Scaling and fitting use training sources only, and held-out predictive R 2 is compared with both within-block and source- by-position-preserving order nulls. Full estima- tor details and results are reported in the Sup- plementary Material because these continuous quantities are related to, but are not the same es- timands as, discrete type accumulation and cat- egorical recurrence. G. Source, acquisition, and conditioned-order controls a. Between- and within-source variation. Agreement at the corpus level does not show whether the measured variation lies between recording sources or among events from the same source. We therefore use two source-aware tests. The first measures how strongly each saved recorded-signal measurement set differs among source blocks with null-adjusted PER- MANOVA R 2 on exact 1,501-event samples. The narrow âsimple acousticsâ set contains duration, spectral centroid, and bandwidth; the archived channel, padding, and all-descriptor sets are retained for the source-localization audit. Larger values indicate greater separation of the measured properties among source blocks after null adjustment. The second removes source-block means from the acoustic measurements and asks whether cluster membership remains associated with dif- ferences among events from the same source. Assignment permutations are restricted within source blocks. For both tests, the saved clus- ter assignments weight encoder families equally and are formed without using sequence order; K = 32 is primary and K = 16, 24 are sensitiv- ity checks. Each cell uses 500 permutations, and results are averaged across 50 alternative cluster constructions (Supplementary Material, Acous- tic variation across source blocks and within- source cluster association). b. Cluster distribution across source blocks. A separate source-block test asks whether the same cluster assignments are distributed across na- tive source blocks more unevenly than expected. The assignments are reallocated across blocks while preserving the exact block sizes. We then compare between-block JensenâShannon diver- gence and the number of distinct clusters within blocks with the corresponding reallocation base- lines (Supplementary Material, Distribution of distinct clusters across source blocks). c. Source- and position-conditioned order nulls. Within-stream position can itself carry stereo- typed sequence structure. We therefore re- peat both immediate dependence and repeated- subsequence recurrence against a stricter null that preserves native source identity and normal- ized stream position simultaneously. Acoustic- cluster assignments are permuted only within native-source Ă position-bin strata, preserving event slots, matched sequence boundaries, se- quence lengths, global occupancy, source-specific inventories, and position-bin cluster counts. Eight position bins with K = 32 are primary; 4 and 16 bins and K = 16, 24 are sensitivity condi- tions, with 500 restricted permutations per cell. For first-order dependence, the reported quan- tity is observed MillerâMadow-corrected entropy reduction minus the conditioned-null mean. For recurrence, compressor size is divided by the conditioned-null mean; ratios below one indicate repeated order beyond source identity and posi- tional composition. J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 5 d. Short-event boundary sensitivity. Events shorter than the 0.175-s encoder window require padding. We record padding fraction and rerun the complete panel with symmetric-reflect rather than right-zero padding to test whether the main repertoire profile depends on this corpus-specific boundary treatment (Supple- mentary Material, Measurement sensitivities and acoustic source analysis). H. Longer predictive context and same-signal timing validation a. Longer predictive context. The primary local-order measurement uses only the imme- diately preceding event. To ask whether one additional event of history contributes predic- tive information, held-out models compare one- event with two-event context. Acoustic clusters are reconstructed separately inside each train- ing split so that sequence outcomes cannot in- fluence how events are grouped. Within each split, encoder vectors are standardized, reduced by principal components, and clustered indepen- dently. The encoder-specific clusterings are then combined according to how often pairs of events are grouped together, with equal weight for each encoder family (Strehl and Ghosh, 2002). This produces a single K = 32 grouping. All prepro- cessing, clustering, combination of clusterings, and assignment of held-out events use training source blocks only. Two transition models score the same held- out events: an order-1 model uses the im- mediately preceding event, and an order-2 model uses the two preceding events. When a longer context is sparse, both models fall back recursively to shorter contexts within the matched 113-block profile. The reported ef- fect is order-2 minus order-1 held-out log like- lihood in bits/token. Positive values mean that the event two positions back improves prediction beyond the immediately preceding event. Uncertainty is estimated by resampling the highest available source block and aggregat- ing over 50 alternative train-only cluster con- structions within each bootstrap replicate. The three planned Whale-minus-reference contrasts are Holm-adjusted and accompanied by simulta- neous 95% source-bootstrap confidence intervals. Order 3 tests whether one additional preceding event adds predictive value. Type accumulation One-step context Repeated subseq. Whale Finch Human Marmoset 122 311 444 233 Rank on each primary measurement FIG. 2. Ranks on the three primary measurements (1 = strongest, 4 = weakest). Whale is top-two on all three; Finch leads the two order-sensitive mea- surements. b. Within-coda timing. Whale codas permit an additional physical test that is not available in the same form for the other native event defini- tions: their within-event timing can be changed while the constituent recorded clicks are held fixed. We therefore re-encode the same clicks after three changes to inter-click-interval (ICI) arrangement: local jitter, replacement from the global ICI distribution, and within-coda ICI per- mutation. Each perturbation changes temporal arrangement within the coda without replacing the click waveforms, providing a within-Whale test of whether coda timing contributes to the measured profile. IV. RESULTS A. Matched primary measurements reveal different repertoire profiles The matched comparison yields different pro- files rather than one ordering (Table I; Fig. 2). Whale leads type accumulation (β = 0.124) and is the only repertoire in the top two on all three primary measurements, ranking 1,2,2. Finch ranks 3,1,1, leading shuffle-adjusted immediate dependence (0.737 versus 0.365 Whale) and Re- Pair/LZ recurrence; DEFLATE is nearly tied (0.964 Finch, 0.965 Whale). Marmoset ranks 2,3,3 and Human 4,4,4. The three-dimensional profile, not an aggregate score, is the principal result. 6 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices The profile is stable across analytical choices: Whale is top-two in 98.2% of encoder/K/reference cells and Finch in 90.5%. Source-block bootstrap separately supports the Whale Heaps direction in 22/33 settings versus Finch and 26/33 versus Human and Marmoset, and the Finch immediate-dependence direction in 24/33 WhaleâFinch settings; DEFLATE re- mains unresolved (Supplementary Material). TABLE I. Means over the matched encoder/selection/K panel. Lower compression ratios indicate greater recurrence. MeasurementWhale Finch Human Marmoset HeapsâHerdanβ0.124 0.045 0.0340.050 Raw first-order gain 1.873 1.230 0.1890.733 Shuffle-adjusted gain 0.365 0.737 0.0650.129 RePair ratio0.928 0.862 0.9910.981 LZ ratio0.945 0.914 0.9950.988 DEFLATE ratio0.965 0.964 0.9990.991 B. Physical acoustics recover complementary parts of the profile a. Physical acoustic dimensions. The first question is whether the encoder-based profile is visible in physically interpretable measurements of the recordings. It is, but different acous- tic families recover different parts of the pro- file (Fig. 3). Whole-profile agreement across the 12 repertoire-by-measurement cells is high- est for spectral shape (Ď = 0.923), followed by acquisition-sensitive level measurements (0.879), temporal acoustics (0.860), and spectral loca- tion/extent (0.713). Metadata is much less sim- ilar (Ď = 0.246). Whaleâs type-accumulation placement is recovered almost exactly by tem- poral acoustics (mean rank 1.013) and spectral shape (1.007), whereas Finchâs local-dependence and repeated-order placement is recovered most strongly by spectral location/extent (1.000 and 1.124) and also by spectral shape (1.380 and 1.364). Thus complementary temporal and spec- tral dimensions recover complementary parts of the encoder result rather than one opaque coor- dinate. The strong acquisition-sensitive agree- ment is retained rather than dismissed: these archival signals contain both production-related and recording-related variation. b. Continuous acoustics without K-means. The acoustic-breadth result also appears without as- 0.00.20.40.60.81.0 Spearman Ď vs. encoder 12-cell profile Spectral shape Acquisition-sensitive Temporal acoustics Spectral location/extent Metadata (reference) 0.923 0.879 0.860 0.713 0.246 A. Whole-profile acoustic agreement W type W local W repeat F type F local F repeat Temporal Spectral location/extent Spectral shape Acquisition- sensitive 1.012.001.793.171.001.21 1.902.001.903.801.001.12 1.011.621.673.631.381.36 1.571.811.123.401.191.89 B. Family-specific placement FIG. 3. Physical-acoustic decomposition of the encoder-derived profile. A: whole-profile agreement with the frozen-encoder pattern. B: family-specific mean ranks for Whale and Finch on the three pri- mary measurements (lower is stronger). Metadata is a non-acoustic reference. signing events to discrete types. Over the full continuous coverage curve, Whale has the largest mean AUC in pooled physical-descriptor space (0.308; Human 0.281, Marmoset 0.229, Finch 0.219) and after native-source means are re- moved (0.482; Human 0.459, Marmoset 0.420, Finch 0.343). Whale is not uniformly largest at every reference-set size: at n = 1200, Whale and Human are nearly tied in the pooled anal- ysis (0.196 and 0.195), and Human is slightly larger after source-mean removal (0.330 ver- sus 0.324). The supported claim is therefore broader whole-curve continuous acoustic cover- age for Whale, not dominance at every sampling scale. A separate leave-one-source-out analy- sis predicts the next continuous acoustic vector from the current one. All four repertoires re- main above both order nulls; under the source- by-position-conditioned null excess held-out R 2 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 7 is 0.305 Whale, 0.206 Finch, 0.140 Marmoset, and 0.131 Human. Continuous vector prediction and recurrence of discrete acoustic types are dif- ferent estimands, so this Whale-leading continu- ous result complements rather than contradicts Finchâs categorical order advantage (Supplemen- tary Material). These representation-independent checks es- tablish that the primary profile has direct physical-acoustic counterparts. They do not yet show where the Whale accumulation pattern lies across recording sources or whether source and stream position can account for the categorical order effects. C. Source and position analyses locate the remaining variation Part of Whale type accumulation lies across sources. Under blockwise redistribution, Finch shows the larger between-block divergence ra- tio (3.152 versus 1.758 at K = 32), whereas Whale shows the stronger within-block richness deficit (0.502 versus 0.735), indicating that more of its pooled inventory is distributed across Date Ă Unit blocks. Direct acoustics agree: simple- acoustic source separation is largest for Whale (R 2 = 0.543; Finch 0.071, Human 0.042, Mar- moset 0.239). Yet after source means are re- moved, cluster membership remains associated with within-source simple acoustics in every repertoire (R 2 = 0.163 Whale, 0.633 Finch, 0.459 Human, 0.372 Marmoset). Whale type accumulation is therefore source-structured but not reducible to between-source differences alone (Supplementary Material). Position also matters but does not explain Finchâs order advantage away. Under the stricter source-by-position null, Finch retains the largest source-aware first-order excess and recurrence effect, with Whale second on both (Table IV). Finch exceeds Whale by +0.872 bits/token [0.520, 1.223] in first-order depen- dence; the Finch/Whale recurrence-ratio quo- tient is 0.935 [0.888, 0.989]. At the primary setting, Finch, Whale, and Marmoset beat the restricted recurrence null in every tokenizer con- struction, whereas Humanâs effect is negligible in magnitude despite a narrow source-aware in- terval. Thus Finchâs categorical order is posi- tionally structured but not reducible to source identity or fixed stream position. Short-event padding is highly asymmetric, but replacing right-zero with symmetric-reflect TABLE IV. Source-aware discrete order beyond source and normalized position. Larger first-order excess and smaller recurrence ratio indicate stronger residual order. RepertoireFirst-order excess Recurrence family ratio Finch1.352 [1.033, 1.672] 0.842 [0.816, 0.878] Whale0.480 [0.337, 0.602]0.900 [0.867, 0.938] Marmoset0.182 [0.141, 0.224]0.971 [0.964, 0.978] Human0.017 [0.007, 0.026]0.998 [0.997, 0.999] padding preserves the secondary Whale/Finch ordering (1.419/1.698 versus 1.448/1.682; Sup- plementary Material). With source, position, and boundary effects bounded, we next ask whether extending temporal history changes the comparison. D. Temporal scale changes the comparison; coda timing pro- vides a same-signal validation Extending discrete predictive context by one event shifts the comparison toward Whale. The order-2 minus order-1 increment is more positive for Whale than Finch by +0.169 bits/token (95% simultaneous interval [0.064, 0.274]), Human by +0.172 [0.074, 0.271], and Marmoset by +0.149 [0.050, 0.248]; all three Holm-adjusted compar- isons satisfy p⤠0.00084. Whaleâs own effect is +0.081 bits/token with source-aware interval [-0.006, 0.177] (p = 0.1361), so the supported result is comparative rather than a positive Whale-alone effect. Finch, Human, and Marmoset have mean effects of - 0.087, -0.091, and -0.068 bits/token, with source- aware intervals wholly below zero. The direction persists in 49/50 train-only cluster constructions while all references are negative in all 50; equal weighting of individual encoder configurations also preserves the direction, and order 3 adds no supported gain (Supplementary Material). Finally, codas permit a same-signal physical test: recorded clicks are held fixed while only their ICI arrangement changes. Observed-coda values exceed all three perturbed versions across five readouts for 10/11 encoder configurations (AVES2 excepted), establishing sensitivity of the Whale profile to observed inter-click timing. 8 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices â0.10.00.10.2 Order 2 - order 1 gain (bits/token) Whale Finch Human Marmoset A. Source-aware order-2 effects 0.00.10.20.3 Difference in order-2 effect (bits/token) Whale - Finch Whale - Human Whale - Marmoset p=.0008 p=.0003 p=.0008 B. Planned Whale contrasts FIG. 4. Held-out two-event context. A: order-2 mi- nus order-1 gain with 95% source-bootstrap inter- vals. B: planned Whale-minus-reference contrasts with simultaneous family-wise 95% intervals and Holm-adjusted p-values. V. DISCUSSION A. The repertoires differ in profile rather than along a single ordering Matching event count and local sequence op- portunity reveals that type accumulation and temporal organization do not produce one com- mon ordering. Whale leads type accumula- tion and is the only repertoire in the top two on all three primary measurements. Finch in- stead leads both order-sensitive measurements but ranks third in type accumulation. Marmoset is second in type accumulation and third on both order-sensitive measurements, while Human is fourth across the three. The scientific result is this three-dimensional profile, not an aggregate ranking of repertoire structure. This comparison is defined at the native event scale of each corpus: phone for Human, coda for Whale, syllable for Finch, and call for Marmoset. These units need not have equal du- rations or internal complexity to be the relevant source-defined events in their respective corpora. Equal event counts and matched block lengths make the opportunity to observe type accumula- tion and local order comparable without assert- ing that the units are biologically homologous or forcing them into an arbitrary common-duration segmentation. B. Physical acoustics provide representation-independent coun- terparts The encoder-derived profile is not confined to a learned representation. Spectral shape, temporal acoustics, and spectral location/extent each recover substantial portions of the com- plete cross-repertoire pattern, and they do so in complementary ways: Whaleâs type accu- mulation is recovered most strongly by tempo- ral and spectral-shape measurements, whereas Finchâs categorical local dependence and recur- rence are especially prominent under spectral lo- cation/extent and spectral shape. Acquisition- sensitive measurements also recover much of the pattern, so these archival recordings retain ac- quisition as well as production-related varia- tion. The physically interpretable decomposi- tion therefore establishes grounding in multi- ple recorded-signal dimensions without implying nuisance-free biological identification. The continuous analyses strengthen that interpretation without relying on K-means. Whale has the broadest whole-curve continuous acoustic coverage both before and after native- source means are removed, although Human approaches it at the largest reference-set size. Continuous next-event prediction, however, is strongest for Whale rather than Finch. That difference is informative: categorical recurrence asks how strongly discrete acoustic types repeat or constrain one another, whereas continuous prediction asks how predictable the next physi- cal acoustic vector is. The two need not produce the same repertoire ordering, reinforcing the pa- perâs central point that temporal organization is multidimensional. C. Source and position refine, but do not erase, the primary profile Source structure still matters. Whale has strong between-source acoustic heterogeneity J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 9 and fewer pooled clusters within its native blocks than expected under redistribution, so part of the pooled Whale type accumulation is dis- tributed across Date Ă Unit source blocks in the sampled corpus of three social units. At the same time, cluster membership remains as- sociated with acoustic differences among events within the same source block. The Whale result is therefore source-structured but not reducible to between-source differences alone. The Finch position result is likewise refined rather than dismissed. A position-only descrip- tion can generate a Finch-like ordering, showing that sequence position carries real structure. Yet when acoustic-cluster assignments are random- ized only among events from the same source and normalized position stratum, Finch retains the largest source-aware excess first-order depen- dence and the strongest repeated-order recur- rence, with Whale second on both conditioned analyses. Finchâs categorical order is therefore positionally structured but not explained away by source identity or a fixed positional template. Short-event padding remains highly asymmetric across corpora, but changing the padding con- vention preserves the main profile. D. Temporal scale changes which repertoire stands out The discrete local-order measurements, con- tinuous next-event prediction, and held-out longer-context test distinguish different aspects of temporal organization. Finch shows the strongest immediate dependence and subse- quence recurrence of discrete acoustic types even after source and position are preserved in the null, whereas Whale shows the strongest contin- uous next-event acoustic prediction. Extending the discrete predictive context by one additional event also shifts the comparison toward Whale: the order-2 increment is more positive for Whale than for each reference repertoire. Whaleâs stan- dalone interval includes zero and order 3 adds no supported gain, so the identified higher-order result is comparative rather than evidence for a positive order-2 effect within Whale alone. A final same-signal intervention links the Whale profile to coda timing. Holding the recorded clicks fixed while changing only within- coda timing makes the observed-coda values ex- ceed all three perturbed versions across the five perturbation readouts for 10 of 11 encoder con- figurations. This physical validation is distinct from the order-2 comparison: it shows timing sensitivity within Whale and is not available in the same form for the other native event defini- tions used here. E. Scope and next tests The results compare four recorded corpora at their source-defined event scales; they do not identify a species-wide ordering indepen- dent of acquisition. Within-repertoire analy- ses spanning smaller and larger source-defined units can test how type accumulation and se- quence dependence change with observation scale. Acquisition-matched recordings with con- trolled microphones, sourceâreceiver distances, environments, and acquisition systems can test whether the observed source-block concentration and cross-corpus profile persist when recording heterogeneity is experimentally constrained. VI. CONCLUSION Under matched event counts and local se- quence opportunities, the four repertoires oc- cupy different acoustic profiles rather than one hierarchy. Whale shows the fastest discrete type accumulation and broadest whole-curve contin- uous acoustic coverage, whereas Finch shows the strongest categorical immediate dependence and repeated-subsequence recurrence even after source and normalized position are preserved in the null. Direct temporal and spectral measure- ments recover complementary parts of this pro- file, while source analyses locate part of Whaleâs accumulation across recording blocks. Temporal organization itself is non- equivalent across representations and scales: continuous next-event prediction favors Whale, extending discrete history shifts the compara- tive effect toward Whale, and same-click ICI perturbations show a physical timing contri- bution within codas. The central result is therefore a separation among acoustic-event diversity, categorical order, continuous acoustic predictability, and longer-context effects. SUPPLEMENTARY MATERIAL See supplementary material at [URL will be inserted by AIP] for estimator definitions, full sensitivity and source-aware results, continuous 10 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices analyses, conditioned-order nulls, higher-order analyses, and Whale timing controls. ACKNOWLEDGMENTS We thank the maintainers of the source cor- pora. This research received no external funding. AUTHOR DECLARATIONS Conflict of interest. The authors have no conflicts to disclose. Ethics approval. This secondary analy- sis uses publicly released recordings and meta- data; approvals for the original collections are described by the source publications. Use of AI-assisted tools. An AI-assisted language tool was used for editorial restructuring and language refinement. The authors reviewed and verified all scientific content, results, cita- tions, interpretations, and final wording. DATA AVAILABILITY Source corpora are available from their origi- nal distributors. Derived manifests, results, and analysis scripts are available during peer review and will be archived publicly before publication; audio and embeddings remain subject to source licenses. REFERENCES Baevski, A., Zhou, H., Mohamed, A., and Auli, M. (2020). âwav2vec 2.0: A framework for self-supervised learning of speech representations,â in Advances in Neural Information Processing Systems, Vol. 33, p. 12449â12460. Bharadwaj, S., Cornell, S., Choi, K., Fukayama, S., Shim, H.-j., Deshmukh, S., and Watanabe, S. (2025). âOpenbeats: A fully open-source general-purpose audio encoderâ arXiv:2507.14129. Deutsch, L. P. (1996). âDeflate compressed data format specification version 1.3,â RFC 1951, doi: 10.17487/ RFC1951. Flores Garcia, H., Seetharaman, P., Kumar, R., and Pardo, B. (2023). âVampnet: Music generation via masked acoustic token modeling,â in Proceedings of the 24th International Society for Music Information Re- trieval Conference, p. 359â366. Gero, S., Whitehead, H., and Rendell, L. (2016). âIndi- vidual, unit and vocal clan level identity cues in sperm whale codas,â Royal Society Open Science 3(1), 150372, doi: 10.1098/rsos.150372. Hagiwara, M. (2023). âAves: Animal vocalization en- coder based on self-supervision,â in IEEE International Conference on Acoustics, Speech and Signal Processing, p. 1â5, doi: 10.1109/ICASSP49357.2023.10095642. Heaps, H. S. (1978). Information Retrieval: Computa- tional and Theoretical Aspects (Academic Press, New York). Hsu, W.-N., Bolte, B., Tsai, Y.-H. H., Lakhotia, K., Salakhutdinov, R., and Mohamed, A. (2021). âHu- bert: Self-supervised speech representation learning by masked prediction of hidden units,â IEEE/ACM Trans- actions on Audio, Speech, and Language Processing 29, 3451â3460, doi: 10.1109/TASLP.2021.3122291. Kershenbaum, A., Blumstein, D. T., Roch, M. A., Akcay, C., Backus, G., Bee, M. A., Bohn, K., Cao, Y., Carter, G., Cesar, C., Coen, M., DeRuiter, S. L., Doyle, L., Edelman, S., Ferrer-i Cancho, R. et al. (2016). âAcous- tic sequences in non-human animals: A tutorial review and prospectus,â Biological Reviews 91(1), 13â52, doi: 10.1111/brv.12160. Koumura, T., and Okanoya, K. (2016). âAutomatic recognition of element classes and boundaries in the birdsong with variable sequences,â PLoS ONE 11(7), e0159188, doi: 10.1371/journal.pone.0159188. Lamothe, C., Obliger, M., Best, P., Trapeau, R., Ravel, S., Artieres, T., Marxer, R., and Belin, P. (2025). âMarmaudio: A large annotated dataset of vocaliza- tions by common marmosetsâ Zenodo, doi: 10.5281/ zenodo.15017207. Larsson, N. J., and Moffat, A. (2000). âOff-line dictionary-based compression,â Proceedings of the IEEE 88(11), 1722â1732, doi: 10.1109/5.892708. McCowan, B., Hanser, S. F., and Doyle, L. R. (1999). âQuantitative tools for comparing animal communica- tion systems: Information theory applied to bottlenose dolphin whistle repertoires,â Animal Behaviour 57(2), 409â419, doi: 10.1006/anbe.1998.1000. Miller, G. A. (1955). âNote on the bias of information es- timates,â in Information Theory in Psychology: Prob- lems and Methods I-B, edited by H. Quastler (Free Press, Glencoe, Illinois), p. 95â100. Morfi, V., Lachlan, R. F., and Stowell, D. (2021). âDeep perceptual embeddings for unlabelled animal sound events,â Journal of the Acoustical Society of America 150(1), 2â11, doi: 10.1121/10.0005475. Morita, T., Koda, H., Okanoya, K., and Tachibana, R. O. (2021). âMeasuring context dependency in birdsong using artificial neural networks,â PLoS Computational Biology 17(12), e1009707, doi: 10.1371/journal. pcbi.1009707. Pitt, M. A., Dilley, L., Johnson, K., Kiesling, S., Ray- mond, W., Hume, E., and Fosler-Lussier, E. (2007). âBuckeye corpus of conversational speech, second re- leaseâ Department of Psychology, Ohio State Univer- sity. J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices 11 Sainburg, T., Theilman, B., Thielk, M., and Gentner, T. Q. (2019). âParallels in the sequential organization of birdsong and human speech,â Nature Communica- tions 10, 3636, doi: 10.1038/s41467-019-11605-y. Sainburg, T., Thielk, M., and Gentner, T. Q. (2020). âFinding, visualizing, and quantifying latent structure across diverse animal vocal repertoires,â PLoS Com- putational Biology 16(10), e1008228, doi: 10.1371/ journal.pcbi.1008228. Sharma, P., Gero, S., Payne, R., Gruber, D. F., Rus, D., Torralba, A., Andreas, J. et al. (2024). âContex- tual and combinatorial structure in sperm whale vo- calisations,â Nature Communications 15, 3617, doi: 10.1038/s41467-024-47221-8. Strehl, A., and Ghosh, J. (2002). âCluster ensemblesâ a knowledge reuse framework for combining multiple partitions,â Journal of Machine Learning Research 3, 583â617. Suzuki, R., Buck, J. R., and Tyack, P. L. (2006). âInfor- mation entropy of humpback whale songs,â Journal of the Acoustical Society of America 119(3), 1849â1866, doi: 10.1121/1.2161827. Watkins, W. A., and Schevill, W. E. (1977). âSperm whale codas,â Journal of the Acoustical Society of America 62(6), 1485â1490, doi: 10.1121/1.381678. Ziv, J., and Lempel, A. (1977). âA universal algorithm for sequential data compression,â IEEE Transactions on Information Theory 23(3), 337â343, doi: 10.1109/ TIT.1977.1055714. 12 J. Acoust. Soc. Am. / 16 September 2026Structure Across Voices