Animal sound classifiers do well at telling species apart. They are much weaker at the finer distinctions field biologists need, such as which individual is calling or which call type in a repertoire is being produced. Current bioacoustic encoders are trained and evaluated on fixed 5- or 10-second crops. Those crops cut through calls and song bouts at arbitrary points. They also discard the structure that most plausibly carries individual and repertoire identity: the order of calls, the gaps between them, and how a phrase is shaped over its length. We propose to train these encoders with a temporal curriculum, meaning a schedule that starts on short inputs and gradually lengthens them. The new element is that each training stage is defined by whole acoustic units, detected the way speech researchers detect prosodic boundaries (the pauses, pitch resets and energy shifts that group speech into phrases), rather than by fixed token counts. The model first sees single calls, then complete bouts, then multi-minute sequences.
The main use of taxonomic resolution here is as the measuring stick. If the idea is right, the gains should concentrate at the fine end of the label hierarchy (individual identity, call type) and be close to zero for coarse species labels. A proposal that improves everything uniformly would just be better pretraining. A proposal whose gains grow as label granularity gets finer would show that sequence structure within vocal units is where the missing resolution lives.
Core Claim
Fine-grained bioacoustic labels depend on sequential structure that spans whole vocal units. A length curriculum whose stage boundaries coincide with those units should therefore produce encoders that beat, at matched compute, both of the following on individual-identification and call-type tasks:
- fixed-crop training;
- a length-matched curriculum that cuts at random positions.
The same comparison should show no meaningful difference on species classification.
The random-cut control is what separates this from existing progressive-length work. GrowLength and HyenaDNA both grow context by doubling token counts on a fixed schedule. GrowLength does this for language-model pretraining, HyenaDNA for DNA up to a million nucleotides. Neither ties stage boundaries to structure in the signal, and the file notes that their schedules are hand-set heuristics. Audio Flamingo 2 is the closest audio precedent. It extends an audio-language model from clips under 30 seconds to 5 minutes in a staged curriculum, but it works on environmental sound and music, uses fixed sliding windows, and is not evaluated on fine taxonomic distinctions.
On the bioacoustic side, the evidence that motivates the gap comes from two papers:
- What Matters for Bioacoustic Encoding trains on 18,000 hours of fixed crops and introduces individual-ID and call-type benchmarks.
- Foundation Models for Bioacoustics - a Comparative Review finds that self-supervised models such as BEATs and BirdMAE only reveal fine temporal-spectral information under attentive probing, a learned attention-based read-out over frame features.
Neither paper varies the temporal extent of training inputs.
Research Program
Unit detector. The first component is a vocal-unit boundary detector that runs without labels on long recordings. Speech prosody research constrains its design. Zimmerman et al. show that pause duration alone misclassifies phrasing, and speakers also rely on pitch and energy shifts. Li et al. show that the subtlest Mandarin boundaries, the minor phrase breaks, need an explicit pitch-and-energy stream fused with a learned spectrogram stream by cross-attention. Their dual-encoder model gained about 20% F1 over single-encoder models on those boundaries. We will build two detectors:
- a cheap one, using energy and spectral-flux thresholds with a silence-gap criterion;
- a learned one, adapting the Li et al. dual-stream design to animal audio, with self-supervised boundary targets from the cheap detector refined by pitch-track discontinuities.
The simpler detector will be used unless the learned one measurably improves downstream results.
Curriculum. The second component is a three-stage curriculum for continued self-supervised training of a BEATs-style encoder, with rotary position embeddings (position encodings that rotate query and key vectors, used by LongRoPE to extend context) so that the model can extend across stages. The stages are:
- Single units. Each input is one detected call, padded to a common length.
- Bouts. Each input is a run of consecutive units ending at the next long gap.
- Long sequences. Each input is a multi-minute stretch.
Recordings come from the long-form sources already used by Miron et al., namely Xeno-canto, iNaturalist and Watkins.
Speech validation. The third component checks the premise on human speech, where gold hierarchical boundaries exist. On the joint Mandarin corpus of Li et al., which annotates four boundary levels (#1 to #4), we train the same curriculum under three kinds of stage boundaries:
- gold boundaries;
- detected boundaries;
- random cuts.
We then measure minor-phrase (#2) detection. This speech experiment isolates whether aligning training units to real structure matters at all, independent of any bioacoustic label noise.
Baselines
Naive reference. The "do nothing" point is the released sl-BEATs-all encoder from What Matters for Bioacoustic Encoding, evaluated as-is with the same attentive probe on every task. Its authors report that it and EffNetB0-all lead most of their 26 datasets. However, no reported result was found on Lacuna for its individual-ID or call-type splits, so we will measure these numbers ourselves. For coarse labels, the file reports only a robustness figure from Miron et al. on BirdSet: self-supervised models lost 0.01 ROC AUC under focal-to-soundscape shift, against 0.09 for supervised ones. No absolute BirdSet or BEANS numbers were found on Lacuna.
Strongest existing method. The strongest competitor is the same encoder given continued self-supervised training for the same number of GPU-hours on fixed 10-second crops from the identical long-recording pool. This matches data, compute and learning-rate search budget. It is strongest because it captures any benefit from simply seeing more in-domain audio.
Stronger length baseline. A GrowLength-style doubling curriculum on the same data tests whether plain length growth, without unit alignment, already delivers the gain.
Key ablation. The random-cut curriculum reproduces the per-stage length distribution of the unit-aligned curriculum exactly but places cuts uniformly at random. This isolates whether unit alignment adds anything beyond the lengths it happens to produce.
Success criterion. Beating the baselines means a gain over both the fixed-crop and random-cut runs on individual-ID and call-type macro-F1 that exceeds seed-to-seed variation across three seeds. That gain must also exceed the corresponding change in species accuracy, meaning a positive interaction between curriculum and label granularity.
Why not just…?
Why not just pool short-crop embeddings over the whole recording? An attentive pooling head over a sequence of frozen 5-second embeddings can see minutes of context without any retraining. This is the most serious threat. The review's finding that attentive probing recovers much of the fine information (Foundation Models for Bioacoustics) suggests the read-out may do most of the work. We therefore include it as a baseline with the same probe budget and the same total context at test time. If it matches the curriculum-trained encoder, the conclusion is that the encoder does not need long training context, only a long read-out, and we will report that.
Why not just train on long inputs from the start? The concern is that fixed long crops are as good as a curriculum when compute is sufficient. HyenaDNA reports instability without a length warm-up and about 40% faster training with one. GrowLength reports lower perplexity at long windows than fixed-length training. Both results come from DNA and text, not audio, so we run fixed long-crop training at matched compute rather than assume the answer.
Why not just segment with an energy threshold and skip the learned detector? This is a legitimate simplification, and the program adopts it by default. Speech evidence suggests thresholds miss the subtle boundaries (Zimmerman et al.; Li et al.). Whether that matters for animal calls is an open question that the curriculum comparison with each detector will answer.
Proposed Experiments and Decisive Checks
Bioacoustic evaluation. All encoders are evaluated with an identical attentive probe on the following:
- the individual-identification and vocal-repertoire benchmarks among the eight new datasets in the Miron et al. 26-dataset suite;
- BEANS, covering birds, bats, dogs, frogs and insects;
- BirdSet, which provides the coarse-label control.
To prevent leakage, individual-ID splits are made by recording session, not by clip.
Decisive check 1: granularity interaction. The unit-aligned curriculum's advantage over random cuts should be significantly positive on individual and call-type tasks and near zero on species tasks.
Decisive check 2: reliance on order. Shuffling the order of units within each test sequence should erase the fine-grained gain but leave species accuracy intact. This would show the model uses sequential structure rather than extra exposure to the data.
Decisive check 3: speech premise. On the Mandarin corpus, gold-boundary stages should beat random-cut stages on #2 detection, with detected boundaries falling between the two.
Failure conditions. The claim fails if any of the following holds:
- the pooling baseline matches the curriculum;
- random cuts match unit-aligned cuts;
- the gains are uniform across granularity levels.
Each of these outcomes would also be informative for how future bioacoustic encoders are trained.
References
- GrowLength: Accelerating LLMs Pretraining by Progressively Growing Training Length
- HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution
- Audio Flamingo 2: An Audio-Language Model with Long-Audio Understanding and Expert Reasoning Abilities
- AVEX: What Matters for Animal Vocalization Encoding
- Foundation models for bioacoustics - A comparative review
- The ICSI+ multilingual sentence segmentation system
- Dual Audio Encoders Based Mandarin Prosodic Boundary Prediction by Using Multi-Granularity Prosodic Representations
- LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens