Approach

We turn raw audio + strong labels (onset, offset, class ∈ {b, mb, h}) into frame-level targets, train a compact CRNN on log-mel features, then post-process frame probabilities into events with thresholding / gap merge and min-duration rules.

---

1) 10-second segments → features & label spans

We work on 10-second analysis windows for quick visualization/debugging. Audio is high-pass filtered (60 Hz), transformed to log-mel (default 128×T), and labels are drawn as a Gantt track.

Waveform (top), standardized log-mel spectrogram (right), and label Gantt (bottom).

Notes

• Sample rate: 16 kHz (configurable).
• Mel params: n_mels=128, n_fft=1024, hop_length=320 (20 ms); configurable in configs/spotter.yaml.
• Frame time mapping: t = frame_index * hop_length / sr.

---

2) Collapsing to a binary track (BS vs NULL)

preparation step for downstream classification, we collapse multi-class labels into a single BS vs NULL track.
Using a 50% overlap rule per frame (any bowel sound type present → BS=1), this simplifies the segmentation task and provides clean BS spans that can then be passed to a classifier to distinguish between {b, mb, h}.

Green bars = frames labeled as “BS”.

---

3) Frame probabilities + threshold

The CRNN outputs per-class frame probabilities. We inspect the probability track and apply a decision threshold .

Dashed line = decision threshold.

Post-processing (configurable)

• threshold: float or list per class
• min_gap_s: merges two same-class islands if the silent gap is shorter than this
• min_dur_s: removes very short events

These parameters can also be learned or optimized via statistical analysis on larger BS datasets or guided by prior research.

These map probabilities → binary masks → merged events with clean onsets/offsets.

---

5) Training curve snapshot

Epoch	Train Loss	Val Loss	FW-F1	Precision	Recall
01	0.5101	0.4563	0.242	0.919	0.139
02	0.4002	0.3581	0.727	0.783	0.679
03	0.3776	0.5387	0.670	0.524	0.929
…	…	…	…	…	…
09	0.3071	0.3039	0.772	0.846	0.710
10	0.2884	0.2905	0.810	0.776	0.847

FW-F1 = framewise F1 computed per frame (hop ≈ 10 ms, window ≈ 25 ms), so temporal resolution is ~10 ms ..

---

4) Predicted events vs Ground Truth

We visualize predicted spans against GT to see timing agreement, merges/splits, and misses at a glance.

Top = GT spans, bottom = predicted spans after post-processing.

---

Key knobs in configs/spotter.yaml:

features:
  n_mels: 128
  n_fft: 1024
  hop_length: 320
postprocess:
  threshold: 0.5
  min_dur_s: 0.08
  min_gap_s: 0.05

t-SNE visualization

This plot (from analysis/explore_tsne.py) shows good class separation .
Separation comes from combining temporal stats and spectral descriptors that capture both "burst shape" and frequency content, giving the features strong discriminative power.

---

---

Sound Event Detection Evaluation (sed_eval.py)

sed_eval.py implements the event-based evaluation metrics described in
de Benito-Gorrón et al. (2021) —

An Analysis of Sound Event Detection under Acoustic Degradation Using Multi-Resolution Systems,
Applied Sciences, 11(23), 11561.
https://doi.org/10.3390/app112311561