Detecting wildlife calls from raw rainforest audio.
Echo is an end-to-end audio classification pipeline for Capuchinbird field recordings. It standardizes raw audio, transforms fixed three-second windows into magnitude spectrograms, trains a compact CNN to recognize call evidence, then scans long recordings and collapses consecutive positive windows into per-file call counts.
STFT turns each audio window into a two-dimensional magnitude map for the CNN.
Fixed inference contract
Every example becomes 16 kHz mono audio with exactly 48,000 samples.
Precision + recall
Both metrics expose performance on imbalanced call and no-call windows.
Postprocessed counting
Adjacent positive windows collapse before the per-recording count is written.
Pipeline
Inference pathOne fixed signal contract from file to count
01Raw audioWAV labels + MP3 field audio
0216 kHz monodecode · mix · resample
033-second window48,000 samples
04STFT320 frame · 32 step
05Spectrogrammagnitude map
06Small CNNbinary probability
07Sliding inferencefixed windows
08Collapse runsadjacent 1s → one event
09Call countper recording
Signal preparationRepresentation + modelCounting logic
Engineering decisions
Decision
Why
Tradeoff
Classify spectrograms instead of raw waveforms
A magnitude STFT gives a compact time-frequency image that a small 2D CNN can learn from.
Frame length and step become fixed preprocessing choices.
Use one three-second input contract
Training clips and inference windows reach the model with the same 48,000-sample shape.
Longer context is discarded and shorter clips require zero-padding.
Collapse consecutive positives
A sustained call can trigger adjacent windows, so runs of positives count once.
Non-overlapping windows can still split a call near a boundary.
Under the hood
Preprocessing
Decode WAV labels and MP3 field recordings
Mix down to mono and resample to 16 kHz
Pad or truncate to 48,000 samples
STFT: 320-sample frame, 32-sample step
Model
Two Conv2D layers with 16 filters
Flatten into a 64-unit dense layer
Sigmoid output for binary classification
Adam + binary cross-entropy
Inference
Scan each recording in fixed windows
Threshold model scores at 0.99
Collapse each consecutive positive run
Export one count per recording
Technical deep dives
STFT
Deep dive A
Audio becomes a learnable image
Echo converts a three-second waveform into a magnitude spectrogram. Time runs horizontally, frequency vertically, and brighter regions carry more energy for the convolutional filters.
A call may remain audible across adjacent windows. Grouping consecutive predictions converts each positive run into one event before summing, so six raw positives above become three calls.
final count = positive runs, not positive windows
Verified repository output
252 calls across 61 of 100 recordings
The checked-in results CSV provides the per-recording counts. This is pipeline output from the supplied forest recordings, not a production quality benchmark.
Evaluation boundaries
Validation uses one held-out notebook split; there is no cross-validation or formal test set.
The notebook evaluates at 0.5 but applies an empirical 0.99 threshold during field inference.
Inference windows do not overlap, so boundary-straddling calls may be missed or split.
The notebook trains in-session and does not persist a model artifact for repeatable deployment.
System checks
Training and inference share the same 16 kHz mono contract.
Every model input is exactly 48,000 samples.
Precision and recall accompany the loss during training.
Inference uses the notebook's explicit 0.99 threshold.
Consecutive positive windows count as one call.
The CSV result contains all 100 field recordings.
Next
Persist the trained model and preprocessing contract, add overlapping inference with event-level evaluation, and compare thresholds through cross-validated precision-recall curves on labeled field recordings.