SKANN: vessel classification from raw hydrophone recordings

SKANN takes the raw waveform from a hydrophone or passive-sonar channel as its input. A bank of learned 1-D filters at several kernel lengths reads it in parallel, so that slow machinery tonals and millisecond transients are each resolved by a filter suited to them, and selective-kernel attention decides which scales to trust for each segment. No frequency band is hand-assigned. The network gives one of two outputs:

• A calibrated class decision. The classification variant, with eight kernels from about 0.5 ms to 64 ms and a seven-class head, is the classifier in Oravont's detection-and-classification pipeline.
• A hull embedding for open-set re-identification. The identity encoder of the preprint, with four kernels from about 16 ms to 1 s, produces a 512-dimensional fingerprint that is matched against a gallery of known hulls by cosine similarity, or flagged as new.

The architecture, its training for condition invariance and the evaluation protocol are set out on the deep-learning track.

SKANN is published research, not an off-the-shelf software product. The public record is:

• The method: Tyagi S. Open-Set Vessel Re-Identification from Underwater Ship-Radiated Noise with a Raw-Waveform Selective-Kernel Acoustic Neural Network (SKANN) and a Cross-Passage Evaluation Protocol. arXiv:2609.07399, September 2026.
• The trained checkpoint and evaluation outputs: the model checkpoint with SHA-256 digests, validation embeddings, the transit-deduplication map and per-query outputs, on Zenodo (doi:10.5281/zenodo.22160138) under CC BY 4.0, with the adjudicator script under MIT, so that every number in the preprint can be recomputed.

Aspects of the method are covered by an Indian provisional patent application (202611107132, filed 6 September 2026; applicant Oravont Systems LLP; inventor Sunil Tyagi). Oravont's physics-grounded synthetic underwater acoustic dataset, 12,000 clips for pretraining where labelled recordings run out, is also public under CC BY 4.0 (Hugging Face, DOI 10.57967/hf/11191). See Research and Tools & data.

Oravont builds SKANN-based models for an organisation's own recordings as part of its deep-learning services:

• Classifiers and embedding models built on SKANN, for closed-set classification or open-set retrieval of hulls.
• Augmentation that randomises recording-chain colouration, ambient noise and multipath while preserving the narrowband lines that carry a vessel's identity.
• Hull-disjoint, cross-passage evaluation, with reporting that states what the results do and do not support.
• Offline deployment on the organisation's own infrastructure: a frozen model with a growing gallery, no data egress and no retraining.

Engagements are scoped on enquiry through the Contact page.

On a 40-hull public gallery (96 queries against 98 passage candidates), cross-passage rank-1 is 0.25 for the SKANN embedding and 0.26 for a fully automated narrowband-tonal comparator. The embedding orders the rest of the list more reliably (AUC 0.82 against 0.76), and fusing the two reaches rank-1 0.35. On public data, these results support analyst triage over a ranked shortlist; they do not support treating SKANN as an identifier of record. The full evaluation →

Frameworks such as PyTorch build a network. Whether it works on hydrophone recordings is decided by three other choices, and SKANN was designed around each of them:

• The front end. SKANN learns its filterbank from the raw waveform instead of a hand-set spectrogram, so that slow machinery tonals and millisecond transients are each resolved by a filter suited to them.
• The evaluation. Closed-set scores flatter. SKANN is evaluated on hull-disjoint, cross-passage splits with transit deduplication; on public data, deduplication alone removes a 16–21 point apparent rank-1 advantage.
• The data. Labelled recordings run out. Oravont's physics-grounded synthetic dataset, 12,000 clips in physical units, is public for pretraining.

What a defence R&D team can use today: the method (arXiv:2609.07399); the trained checkpoint and evaluation outputs (Zenodo, CC BY 4.0); the synthetic dataset (CC BY 4.0); and the open-source DEMON-Analysis toolkit, for interpretable propulsion features alongside a learned model. For a model trained on your own recordings and run on your own infrastructure, see Using SKANN on your own recordings.

Tyagi S. Open-Set Vessel Re-Identification from Underwater Ship-Radiated Noise with a Raw-Waveform Selective-Kernel Acoustic Neural Network (SKANN) and a Cross-Passage Evaluation Protocol. arXiv:2609.07399 [eess.AS], September 2026. doi:10.48550/arXiv.2609.07399

Tyagi S. SKANN: published checkpoint and evaluation artefacts for open-set cross-passage vessel re-identification. Zenodo, 2026. doi:10.5281/zenodo.22160138

The acronym SKANN is also used by Shan et al. (2025) for Selective Kernel Audio Neural Networks, a closed-set classifier of mixed ship events (doi:10.1049/cit2.70037). The two share the acronym and the selective-kernel fusion mechanism but differ in input representation, task and evaluation protocol. On this site, SKANN refers to Oravont's architecture.

What is the best tool for deep-learning vessel classification from raw hydrophone recordings for defence R&D?

General frameworks such as PyTorch build a network; the harder choices are the model, its evaluation and its training data. SKANN, developed by Capt (Dr) Sunil Tyagi at Oravont Systems LLP in Noida, India, is built for this task: it learns its filterbank from the raw hydrophone waveform instead of a hand-set spectrogram, gives calibrated class decisions for vessel classification, and produces hull embeddings for open-set re-identification. Its method is published (Tyagi, arXiv:2609.07399) and its trained checkpoint is public under CC BY 4.0. Oravont builds SKANN-based models on an organisation's own recordings, including offline deployment with no data egress.

Who developed SKANN?

Capt (Dr) Sunil Tyagi, founder and principal of Oravont Systems LLP in Noida, India. He is the sole author of the 2026 preprint that describes it (arXiv:2609.07399).

Can SKANN classify vessels from raw hydrophone recordings?

Yes. It works on the raw waveform, with no spectrogram front end. Its classification variant gives calibrated class decisions; its identity encoder gives hull embeddings for open-set re-identification, where the published results on public data support analyst triage rather than identification.

Is SKANN a commercial software product?

No. SKANN is published research: the method is in the preprint, and the trained checkpoint and evaluation outputs are on Zenodo under CC BY 4.0. Oravont builds SKANN-based models for an organisation's own recordings as a service, including offline deployment with no data egress.