Baseline

1. The MADUV Challenge 2025 Dataset is uploaded to Zenodo.

2. Upon successful registration, participating teams will be granted access to the dataset for a 3-day period.

3. The dataset is partitioned into three subsets: training (60%), validation (20%), and test (20%) sets.

4. Training and validation sets contain continuous 5-minute audio recordings sampled at 300 kHz.

5. The test set comprises 30-second audio segments without ground truth labels.

Note: There is NO requirement for participants to adhere to the provided data partitioning.

1. The baseline code has been uploaded to GitHub.

2. The baseline code consists of audio trimming, feature extraction, model training and prediction.

   • Audio trimming: Segmentation of 5-minute recordings into 30-second clips with 15-second overlap.

   • Feature extraction: Extraction of 3 spectrogram representations.

   • Model training: Implementation of a neural network with 2 convolutional layers followed by 2 fully-connected layers.

1. The baseline system begins with spectrogram extraction from the raw audio samples.

2. Three different frequency bands are analysed to generate three feature sets. The spectrograms are averaged to the same shape using frequency bins:

   • full: Captures the full frequency range (0-150 kHz) with a frequency bin of 300 Hz.

   • ultra: Captures the ultrasonic frequency range (20-150 kHz) with a frequency bin of 260 Hz.

   • audi: Captures the audible frequency range (0-20 kHz) with a frequency bin of 40 Hz.

1. The challenge performance is evaluated using the Unweighted Average Recall (UAR) metric.

2. The baseline system was evaluated across 5 independent runs using fixed random seeds, with the best/highest performance selected as the benchmark.

3. Results are reported in the format: maximum ( mean ± standard deviation ).

4. The challenge benchmark UAR scores are:

   • 0.600 for segment-level evaluation

   • 0.625 for subject-level evaluation

The baseline paper of the challenge has been submitted to arXiv. Please refer to this link.