Lab 7 · PNW-Cnet + Train a Custom Spectrogram CNN

Biological motivation

You are monitoring an old-growth forest where the species of management interest is the barred owl pair duet — a distinctive antiphonal call between a male and a female. PNW-Cnet has a "barred owl" class but it's trained on single-bird calls and treats duets the same as solo calls. You need a classifier that specifically detects the duet to estimate pair occupancy.

Data

You'll use two pieces:

• data_sample/audio/ — the 97-clip iNaturalist sample shipped with the course (~5–30 s each), indexed in data_sample/audio_index.csv by ground-truth species. We'll treat this as a "stand-in soundscape" for the inference task. (A real deployment would have 4–8 h of continuous ARU recordings; for Lab 7's smoke test the per-recording snippets are enough.)
• Custom-CNN training set: the 97 iNat clips (10 per species, except 9 varied thrush and 8 robin) above, grouped into three biologically distinct super-classes that a small CNN can plausibly learn:
  • owl: barred owl + northern spotted owl (20 clips)
  • songbird: thrushes, wren, junco, robin (60 clips)
  • other: pileated woodpecker, olive-sided flycatcher (20)
  In place of the "barred-owl pair duet vs solo" target described above, the deliverable for this lab uses this 3-super-class formulation (no hand-labeling required). If you want to do the duet/solo task instead, ask the instructor for the additional hand-labeled clips.

Tasks

Task 1 — PNW-Cnet inference via pycnet-audio (4 pts)

Install pycnet-audio in its own conda env (see the bioacoustics tutorial):

conda create -n pycnet -c conda-forge sox python=3.11 -y
conda activate pycnet
pip install pycnet-audio

Then run the end-to-end pipeline on the soundscape directory using PNW-Cnet v5:

pycnet process /path/to/soundscapes -c v5 -w 4 -a

This generates 12-second non-overlapping spectrograms, scores them, and writes:

• soundscapes_v5_class_scores.csv — one row per 12-s segment, Filename + 135 class-score columns.
• soundscapes_v5_detection_summary.csv — aggregated counts by date / threshold.
• soundscapes_review_kscope.csv — review file for Kaleidoscope-format human verification.

Convert the class-scores CSV to long form (one row per above-threshold detection) and save as pnwcnet_detections.csv. Report the top 5 detected classes by frequency and note which (if any) are barred owl variants (e.g., STVA_8Note, STVA_Duet).

Task 2 — train custom 3-class CNN (8 pts)

Use train_spec_cnn.py (from the bioacoustics tutorial) as a starting point, adapting for 3 classes (duet, solo_barred, background). Run for at least 10 epochs. Save:

• training_curves.png (train + val loss, train + val accuracy).
• val_confusion.png.
• custom_cnn.pt (best validation epoch).

Task 3 — compare custom CNN to PNW-Cnet on the duet question (4 pts)

For each 5-second window in the soundscape:

• Get PNW-Cnet's barred-owl score for the overlapping segment.
• Get your custom CNN's duet score for the same window.

Plot a scatter of the two scores. Comment on how well PNW-Cnet's general "barred owl" score serves as a proxy for the more specific "duet" event you actually care about.

Task 4 — interpretation (4 pts)

In 200 words: when does the off-the-shelf model suffice, and when do you need to train your own? What does your custom CNN's failure mode look like (listen to a few false positives) and how would you improve it?

Go-further

Augment the duet class by temporally jittering and pitch-shifting clips (± 1 semitone) and re-train. Did augmentation help?

Rubric