Abstract
Maritime distress communications over VHF radio are safety-critical voice messages governed by the Global Maritime Distress and Safety System (GMDSS). In practice, automatic analysis remains difficult because messages are often brief, noisy, and produced under stress, may deviate from the prescribed format, and are further degraded by automatic speech recognition (ASR) errors caused by channel noise and speaker stress.
This paper presents SeaAlert, a controlled experimental framework for evaluating robust analysis of maritime distress communications using transformer-based severity classification and LLM-based structured extraction. To address the scarcity of labeled real-world data, we develop a synthetic data generation pipeline in which an LLM produces diverse maritime messages, including challenging variants in which standard distress codewords are omitted or replaced with less explicit expressions. The generated utterances are synthesized into speech, degraded with simulated VHF noise, and transcribed by an ASR system to obtain controlled noise-degraded transcripts.
The resulting evaluation shows that transformer-based classification (RoBERTa) degrades more gracefully than lexical baselines under ASR noise and codeword masking, while LLM-based extraction (GPT-4o-mini) is more effective than Regex-based extraction for noisy structured fields.
Methodology
SeaAlert addresses two goals: Severity Classification (Routine / Safety / Urgency / Distress) and Information Extraction (Vessel, Call Sign, MMSI, Position, POB, Nature of Incident) from noisy ASR transcripts. A synthetic dataset of maritime radio messages, generated by GPT-4o-mini, is synthesized into speech, degraded with simulated VHF channel noise, and transcribed by Whisper ASR to obtain realistic noisy transcripts.
We compare a rule-based GMDSS keyword spotter, a TF-IDF + Logistic Regression baseline, and a fine-tuned RoBERTa model across four experiments: ASR robustness, adversarial scenarios, codeword-masking ablation, and Regex vs. GPT-4o-mini structured extraction.
Case Study: A Message's Journey Through the Pipeline
From Clean Text to Noisy ASR Transcript, Then Through Classification and Extraction
PAN PAN, PAN PAN, this is MV Ocean Voyager, Call Sign ABC1234, MMSI 123456789. We are currently grounded at position 35 degrees 42.5 minutes North, 139 degrees 30.2 minutes East. Weather conditions are moderate with 10 knots of wind from the east and a slight swell. There are 12 persons on board. We require assistance to refloat our vessel. No injuries reported at this time. Over.
Synthesized via Coqui TTS, then degraded with simulated VHF radio channel noise (SNR: High Noise condition)
PAM PAM. PAM PAM. This is my Diversal Ocean Voyager, Call Sign ABC-134-MMS-I-123,456,789. We are currently grounded at position 35 degrees, 42.5 minutes north, 139 degrees 30.2 minutes east. Weather conditions are moderate with 10 knots of wind from the east and the slight snow. There are 12 persons on board, who require assistance to reflow our vessel. No injuries supported at this time. Over.
Classification Results
| Ground Truth | Logistic Regression Prediction | RoBERTa Prediction |
|---|---|---|
| Urgency | Routine (Conf: 0.39) | Urgency (Conf: 0.92) |
Information Extraction Results
| Field | Ground Truth | GPT-4o-mini Extraction | Status |
|---|---|---|---|
| Vessel | MV OCEAN Voyager | Diversal Ocean Voyager | Partial |
| Call Sign | ABC1234 | ABC-134-MMS-I-123 | Partial |
| MMSI | 123456789 | 123456789 | Exact Match |
| Position | 35° 42.5' N, 139° 30.2' E | 35° 42.5' N, 139° 30.2' E | Exact Match |
| POB | 12 | 12 | Exact Match |
| Incident | Grounding | Currently grounded, no injuries reported | Correct |
Key Findings
| Experiment | Rule-Based GMDSS Spotter | BoW (Logistic Regression) | RoBERTa |
|---|---|---|---|
| Clean Text Accuracy | 57.3% | 68.0% | 67.6% |
| Clean Text Macro-F1 | 0.585 | 0.674 | 0.670 |
| ASR Medium Macro-F1 | — | 0.427 | 0.619 |
| ASR High Macro-F1 | — | 0.423 | 0.595 |
| Relative Drop (Clean → ASR High) | — | ~37% | ~11% |
| ECE at ASR High (lower is better) | — | 0.236 | 0.068 |
| Adversarial Trap Accuracy (n=15) | — | 26.7% | 40.0% |
Takeaway: Clean-text performance is comparable across models, but RoBERTa's real advantage emerges under degradation - a much smaller Macro-F1 drop, better-calibrated confidence, and stronger fallback behavior when GMDSS codewords are masked or removed. For structured extraction, GPT-4o-mini zero-shot surpasses Regex on every field at high ASR noise, especially alphanumeric identifiers (MMSI, Call Sign). Both classifiers still fail on drill-message traps, underscoring the need for human oversight.
BibTeX
@article{Atia2026SeaAlert,
title={SeaAlert: Robust Severity Classification and LLM-Based Information
Extraction for Noisy Maritime Distress Communications},
author={Atia, Tomer and Aperstein, Yehudit and Apartsin, Alexander},
journal={IEEE Access},
year={2026},
doi={10.1109/ACCESS.2026.3709004},
url={https://ieeexplore.ieee.org/document/11592329}
}