Acoustic Detection Model Trainer · Tab two

Learn the acoustic fingerprint of a known vessel.

The Vessel Identity Model Trainer organises recordings by vessel type and MMSI, groups adjacent clips into encounters, visualises representative signatures and trains named models to distinguish individual known vessels.

Analyze · Vessel Identity
Analyze Vessel Identity Model Trainer showing vessel groups, recordings, named profiles and Train or Retrain Model
The Vessel Identity tab mirrors the Sound Source workflow while adding MMSI, encounter and signature-specific tools.

What this trainer answers

“Which known vessel produced this sound?”

This trainer is designed for identity-level recognition. Each class represents a known vessel, normally anchored by MMSI, rather than a broad sound category. Vessel type remains useful for organisation, but the learned prototypes represent individual identities.

Separate model, separate question

The Sound Source trainer can distinguish Vessel from Orca, Sonar or Background. The Vessel Identity trainer operates after the relevant audio is known or suspected to contain a vessel and asks which trained identity is the closest match.

Open-set caution

Real deployments contain vessels absent from the library. An adjustable similarity cutoff allows the model to return Unknown vessel rather than forcing every recording into a known MMSI.

Vessel Library controls

Manage identity-labelled acoustic evidence.

Add Vessel

Create an identity using vessel name, MMSI and vessel type. User recordings are stored under Sound Library/Vessel Identities.

Rename

Change a vessel’s display name without discarding its stable MMSI association.

Add Audio Files

Copy WAV or FLAC recordings into the selected vessel’s managed folder and manifest.

Load in Analyze

Send a selected WAV to the main workspace for waveform, spectrum and spectrogram inspection.

Play / Stop Audio

Audition vessel examples directly and identify poor, silent or contaminated recordings.

Remove safely

User recordings move to a recoverable .excluded folder; downloaded PORTIA references remain read-only.

Selection and encounters

Control exactly which evidence enters a model.

Checkboxes operate at vessel-type, MMSI and individual-file levels. Select All and Deselect All provide the same bulk workflow as the Sound Source tab. The selected file set is passed directly into training, allowing distinct models to use different subsets of the same library.

Encounter grouping

Analyze assigns adjacent clips from one MMSI to a shared encounter using their timestamps. By default, a gap greater than 30 minutes begins a new encounter. Encounter IDs remain visible beside each file.

Leakage detection

The library reports encounters crossing training, validation and test splits. Batch verification automatically excludes any held-out encounter that also contains a training clip.

Read the detailed encounter-grouping guide →

Vessel signature workspace

See and hear what the model is learning from.

Identity table

Review MMSI, type, window totals, train/validation/test counts and whether each identity is currently trainable.

Mean spectrum

Compare channel-specific frequency-energy structure using PyQtGraph and the application colour theme.

Spectrogram

Inspect the representative recording’s time-frequency pattern for tonal machinery and broadband components.

Representative audio

Play a model-representative WAV and connect the numerical signature to the audible recording.

Multi-channel signatures

Preserve available hydrophone channels so channel-specific spectral structure remains visible instead of being collapsed prematurely.

Screen-safe popup

The signature window remains inside the available desktop and keeps controls reachable above the taskbar.

Step-by-step

Train a vessel identity profile.

  1. Add or import vessel recordings. Confirm that every recording belongs to the stated MMSI.
  2. Review encounters. Avoid relying on many adjacent clips from only one passage.
  3. Select files and vessels. Use tree or file checkboxes to define the training subset.
  4. Create or choose a model profile. Named models allow different vessel sets or experiments to coexist.
  5. Set minimum recordings and embedding dimensions. Identities require enough training material plus validation and test examples.
  6. Train or retrain the model. Analyze extracts spectral features, standardises them, learns a reduced discriminative embedding and builds a prototype for each eligible MMSI.
  7. Inspect vessel signatures. Listen to representative audio and examine spectrum and spectrogram evidence.
  8. Run batch verification. Evaluate genuinely held-out encounters and adjust the unknown cutoff using independent evidence.
  9. Activate the trained profile. Retain earlier versions through Model History.

Model and outputs

An interpretable spectral embedding baseline.

StageWhat it doesWhy it matters
Spectral bandsSummarises absolute-frequency energy and short-term level statistics from active hydrophone channels.Captures tonal and broadband machinery structure.
Standardisation and PCAScales features and removes redundant dimensions.Produces a compact representation from high-dimensional spectra.
Linear discriminant embeddingFinds directions that separate the trained MMSIs.Focuses the space on differences among known vessels.
Cosine prototypesStores a normalised centroid for every vessel identity.Supports similarity ranking and representative-file selection.
Unknown cutoffRejects the best candidate when similarity is insufficient.Avoids forcing every deployment sound into a trained identity.

Batch verification

Measure performance by file and by passage.

The verifier reports expected MMSI, predicted identity, cosine similarity, strongest alternative, split, encounter and result status for every available held-out file. It then calculates file accuracy, encounter-majority accuracy and expected-to-predicted confusion counts.

Encounter-majority accuracy is usually the more meaningful summary because adjacent clips within one passage are correlated. Close alternative scores identify vessel pairs that need additional independent data or more varied operating conditions.

Read the complete batch-verification guide →

Decision support: a close acoustic match does not prove identity. AIS association, competing vessels, propagation conditions and analyst review remain important.

Develop vessel identity models from inspectable evidence.

We can demonstrate library organisation, signature review and encounter-aware verification.

Plan a vessel model demonstration