An Open Reproducible Framework for CNN-Based Cetacean Vocalization Detection in Passive Acoustic Monitoring

This paper introduces the open-source, reproducible `ai-pam-pipeline` framework for CNN-based cetacean vocalization detection, demonstrating through controlled experiments that preprocessing choices like FFT window length significantly impact cross-domain generalization while achieving high performance in both binary and multiclass detection tasks.

The Gist

Imagine you are trying to listen to a specific type of bird singing in a very noisy forest, but you can't use your ears; you have to use a computer program to "see" the sound waves on a screen. This paper introduces a new, open-source tool (like a free, shared recipe book) that helps scientists do exactly that for whales and dolphins.

Here is the breakdown of what the paper does, using simple analogies:

1. The "Universal Recipe" (The Framework)

Think of the authors' tool, called ai-pam-pipeline, as a master kitchen. Instead of every scientist building their own stove, oven, and mixing bowls from scratch, they all use this same, pre-built kitchen.

• The Benefit: You just turn a single dial (a configuration file) to change the settings. This means if you cook a dish today and someone else cooks it tomorrow using the same dial settings, they get the exact same result. No more "it worked on my machine" excuses. It works for any type of whale or dolphin, not just one specific kind.

2. The Experiment: How Sharp Should the Lens Be? (Experiment A)

The scientists wanted to know: Does the way we turn sound into pictures matter?

• The Analogy: Imagine taking a photo of a dolphin's whistle. You can take a photo with a low-resolution camera (blurry, big pixels) or a high-resolution camera (sharp, tiny pixels). In this study, they tested three different "camera settings" (called FFT window lengths: 256, 512, and 1024).
• The Result at Home (In-Domain): When they tested the dolphins in the exact same environment where the tool was trained (like taking photos in the same room), all three camera settings worked perfectly. It didn't matter which one they used; the dolphins were easy to spot.
• The Result on the Road (Cross-Domain): When they took the tool to a new environment (a different ocean with different background noise), the results changed dramatically.
  • The "low-resolution" setting (256) was the clear winner.
  • Why? The paper explains this with a cool visual trick. When the computer takes a blurry, low-resolution sound image and stretches it to fit a standard size, the "blurry" parts actually become thicker, brighter, and easier to see. It's like taking a small, fuzzy sketch of a dolphin and blowing it up on a wall; the fuzzy lines become bold, high-contrast shapes that the computer can easily recognize. The sharper settings, when stretched, actually lost some of that helpful contrast.

3. The "Perfect Score" (Thresholds)

The scientists worried that maybe the "low-resolution" setting only looked good because they were cheating by changing the "pass/fail" line (the threshold).

• The Reality Check: They tested every possible pass/fail line from 10% to 90%. The result? The low-resolution setting got a perfect score (1.000 precision) no matter where they set the line. This proves the advantage wasn't a trick; it was a genuine improvement in how the sound looked to the computer.

4. The Hard Part: Sorting the Noise (Experiment B)

The tool isn't just for finding if a dolphin is there; it can also tell you what kind of sound it is making.

• The Challenge: They taught the tool to sort five different types of dolphin sounds. It did a great job overall.
• The Confusion: Sometimes, the tool got confused between two specific sounds: "click trains" and "burst-pulse sounds."
• The Reason: This wasn't because the computer was "stupid." It's because, biologically, these two sounds are so similar to each other that even a human expert might struggle to tell them apart instantly. The tool is actually reflecting the reality of the animal's biology, not a failure of the software.

The Bottom Line

The main takeaway is simple: How you prepare the data matters more than you think.
The paper shows that a small, often-overlooked choice (like how you slice the sound into pieces before analyzing it) can make or break a system when it tries to work in a new environment. By using their open, reproducible framework, scientists can now test these choices systematically to make sure their "whale detectors" work everywhere, not just in the lab.

Technical Summary

Technical Summary: An Open Reproducible Framework for CNN-Based Cetacean Vocalization Detection

Problem Statement
Passive Acoustic Monitoring (PAM) is critical for cetacean research, yet the field often lacks standardized, reproducible workflows for Convolutional Neural Network (CNN)-based detection and classification. A specific gap exists in understanding how preprocessing choices—often treated as secondary implementation details—affect model generalization across different acoustic domains. Furthermore, there is a need for open-source toolkits that allow for systematic parameter evaluation while guaranteeing exact experimental reproducibility.

Methodology
The paper introduces a six-stage methodological framework implemented as the open-source toolkit ai-pam-pipeline. This framework is designed to be generalizable across species and is fully parameterized via a single configuration file, ensuring that experimental conditions can be exactly replicated. The methodology employs CNNs for both binary detection and multiclass classification of cetacean vocalizations.

To validate the framework, the authors conducted two primary experiments:

1. Experiment A (Binary Detection): This study investigated the impact of the Fast Fourier Transform (FFT) window length ($N_{fft}$) on the detection of Bottlenose dolphin (Tursiops truncatus) whistles. The study tested three window lengths: 256, 512, and 1024. Evaluation was performed using stratified 10-fold cross-validation on two datasets: an in-domain dataset (Oltremare, 192 kHz) and a cross-domain benchmark (DCLDE 2022).
2. Experiment B (Multiclass Classification): This experiment demonstrated the framework's capability to classify five distinct T. truncatus vocalization categories.

Key Results

• In-Domain Performance: On the in-domain dataset, performance was uniformly high across all $N_{fft}$ configurations, with a macro F1 score of approximately 0.98. Statistical analysis (Wilcoxon test) showed no significant differences between the window lengths ($p > 0.05$).
• Cross-Domain Performance: Results diverged significantly when applied to the cross-domain benchmark. An $N_{fft}$ of 256 proved significantly superior to larger window lengths ($p = 0.006$, rank-biserial $r = 0.89$).
• Mechanism of Superiority: The authors attribute the superior performance of the smaller window length to an "upsampling amplification effect." Coarser spectral bins (resulting from smaller $N_{fft}$) produce wider, higher-contrast frequency-modulated (FM) traces after the spectrograms are bilinearly resampled to fixed image dimensions for CNN input.
• Threshold Invariance: The advantage of $N_{fft} = 256$ was found to be threshold-invariant. Precision remained at 1.000 across all configurations and decision thresholds ($\theta \in [0.1, 0.9]$), confirming that the performance gain is not an artifact of specific threshold choices.
• Multiclass Capability: In the multiclass experiment, the framework achieved a macro F1 score of 0.843. The analysis noted that inter-class confusion between click trains and burst-pulse sounds reflected biological signal overlap rather than classifier failure.

Significance and Claims
The paper claims that preprocessing choices, frequently overlooked as minor implementation details, can significantly influence cross-domain generalization in PAM tasks. While the study uses $N_{fft}$ as a controlled case study, the primary significance of the work lies in the ai-pam-pipeline framework itself. The authors posit that this toolkit enables the systematic and reproducible evaluation of arbitrary preprocessing parameters within a unified experimental protocol. By providing a fully parameterized, open-source solution, the framework aims to standardize how researchers evaluate and report the effects of methodological variations in cetacean vocalization detection.