Choose one of these questions for your own recordings.

Find the vocalization.

Ignore the recorder.

Keep the timing.

Which strong acoustic events in this figure are not the target vocalizations?

When you run the same checkpoint on a new WAV file, which box is absent?

Move the selected frame. Which parts of the segment could change its interpretation?

Illustrative numbers, not measured checkpoint outputs.

These are learning choices—not filters you turn on while running an existing model.

Name one way your data differ from the data used to train the model you plan to use.

At which stage do we obtain embeddings? At which stage can a threshold change the result?

Inspect false negatives as well as the events the model already found.

A new WAV file enters a fixed checkpoint.
What happens next?

Two minutes in pairs. Identify one remaining uncertainty.

After the break: prepared CoCoMo environment → CLI tour → first inference.

a2v2-infer MODEL.native.pt AUDIO.wav events.tsv \
  --threshold 0.50 --method avg --fusion-window-seconds 0

Raise the threshold: which weak events vanish, and which runs fragment?

a2v2-infer MODEL.native.pt AUDIO.wav events.tsv \
  --threshold 0.50 --method max --fusion-window-seconds 0.10

Average can attenuate short peaks. Maximum can extend or merge nearby events. Inspect both timing and false positives.

project/
  wav/clip01.wav
  lbl/clip01.h5
  manifests/train.tsv
  manifests/valid.tsv

train.tsv · first line: audio root · remaining fields separated by a literal TAB

/datasets/project
wav/clip01.wav	80000

The row means 80,000 waveform samples—not 80,000 model frames. Set task.data to the manifest directory.

# Toy vocabulary: task.unique_labels = ["call_A", "call_B"]
# 8 kHz audio; call_A from 1.00 s up to (not including) 1.15 s
start_frame_lbl = [8000]
end_frame_lbl   = [9200]
lbl_cat         = [0]

Training requires the configured audio rate. min_label_size filters HDF5 file bytes; 3032 is not a minimum number of calls.

What would you need to show before calling a feature “individual identity”?

Arrays are returned as CPU tensors. Exact T comes from the audio geometry and endpoint trimming.

Can these two frame sequences have the same mean but different biological interpretations?

The archive has no dedicated embedding-export CLI. Use the Python API in the practical session.

from a2v2 import InferenceRunner
from a2v2.data import load_audio

runner = InferenceRunner("MODEL.native.pt")
waveform, sample_rate = load_audio("AUDIO.wav")
result = runner.run_tensor(waveform, sample_rate)

Z = result.embeddings.numpy()     # [T, D]
t = result.timestamps.numpy()     # [T], seconds
p = result.probabilities.numpy()  # [T, C]

Use the provided native frame checkpoint. Replace the two illustrative filenames.

Retrieval and exploration can be unlabelled. A labelled downstream classifier is supervised learning.

Choose one unit and one metadata field you will inspect first.

Illustrative schema—not actual participant recordings.

Which level will your result generalize to: another frame, recording, individual, or site?

# Z: [T, D]; t: frame-center timestamps in recording seconds
# start_s and end_s come from a declared annotation/detection policy
inside = (t >= start_s) & (t < end_s)
if not inside.any():
    raise ValueError("No frame centers fall inside this interval")
z_event = Z[inside].mean(axis=0)   # [D]

Very short intervals may contain no frame center: flag them, rather than silently producing NaNs.

Store both the source interval and the extraction segment/context policy.

For your question, could vector magnitude carry useful information—or a nuisance factor?

features
  row_id  → vector

metadata
  row_id  → file, interval,
            individual, site,
            device, session

Keep missing metadata explicitly missing; do not invent individuals, sites, or devices.

analysis_bundle/
  embeddings.npy        # rows identified by rows.csv
  rows.csv              # row_id, source interval, metadata
  extraction.json       # checkpoint and processing provenance

Proposed export convention for the practical—not a built-in CLI output schema.

Reveal the colours by call group, site, and device. Which interpretation survives?

A method comparison is uninterpretable when the underlying observations also change unnoticed.

Pick one query observation. Explain what would count as a useful neighbour.

In your real dataset: inspect the audio, remove self-matches, and avoid near-duplicate segments.

The graph below is a schematic. No UMAP fit is computed in this slide.

Cluster assignments are fitted in the 12-D feature space. The PCA coordinates only display them.

Grey observations are marked noise. “Noise” here is an algorithmic status—not a biological annotation.

What observation would make you reject your current interpretation of a cluster?

Rare vocalization? Corrupted file? Unusual background? Annotation mistake? Listen before naming it.

State a hypothesis with no labels. Then reveal site, device, and call group in that order.

A neighbour is a retrieval result.
A cluster is a candidate grouping.
An outlier is a review priority.

Which analysis will you try first—and what would convince you that it is useful?