The question behind the work.
How can noisy, multi-annotator audio labels be converted into a robust multi-label classification pipeline without leakage between recording splits?
Dataset & environment
University sound-event dataset: 15 event categories, segment-level acoustic features and annotated recordings. The reported segment count should be confirmed against the original files.
From the idea to the method.
- Aggregate annotation signals into segment-level multi-label targets; inspect label prevalence and ambiguous annotations.
- Extract frame/segment-level MFCC acoustic features; define a consistent training and inference representation.
- Split train/validation/test data by recording identity so overlapping segments from a recording do not leak between partitions.
- Train and evaluate an AudioMLP baseline using BCEWithLogitsLoss, appropriate thresholding and per-class diagnostics.
- Compare multi-label classification behavior across classes and report label sparsity, imbalance and limitations.
What was implemented.
- University MLPC project involving sound-event labels and multi-label acoustic classification.
- Design and analysis of annotation aggregation, MFCC-based features and a neural network pipeline.
How the work is evaluated.
Evaluate macro/micro F1, per-class precision/recall, threshold sensitivity and recording-disjoint splits. Exact university experimental scores will be added after original artifacts are linked.
What is documented.
Academic project work has been documented conceptually. Dataset totals and model metrics require final checking against the university project files; no benchmark accuracy is invented here.
Original dataset, code and precise evaluation artifacts should be linked before listing numerical benchmark results.
A foundation for the next experiment.
- In audio recognition, splitting by recordings is critical when many segments originate from the same source.
- Multi-label objectives and threshold selection matter as much as model architecture.