🐒 Sauim Bioacoustic Detector

sauim-detector is a Python command-line tool for bioacoustic processing and automatic detection of Pied tamarin (Saguinus bicolor) vocalizations. It uses a pre-trained bird vocalization embedding model and a custom One-Class SVM classifier trained to detect the target species.

Quick tutorial: Watch on YouTube

Installation

Create and activate a Python 3.12 environment. This step is recommended to keep dependencies isolated; the environment name can be anything:

conda create -n sauim python=3.12 pip
conda activate sauim
python -m pip install --upgrade pip

Install the released package from PyPI:

python -m pip install sauim-detector

For local development, clone the repository and install the package in editable mode:

git clone https://github.com/juancolonna/Sauim.git
cd Sauim
python -m pip install -e ./sauim-detector

The package requires Python >=3.12,<3.13. TensorFlow and TensorFlow Hub compatibility is sensitive to Python and setuptools versions, so the project pins setuptools<82.

Usage

Run the detector with a path to a .wav file:

sauim-detector records/Mindu_Saguinus_bicolor_02.02.19-000.wav

The --stride argument sets the hop length, in seconds, between consecutive analysis windows. It accepts values from 1.0 to 5.0; the default is 5.0. Lower values use more overlap between windows and may improve coverage of short vocalizations, but increase processing time. Example:

sauim-detector records/Mindu_Saguinus_bicolor_02.02.19-000.wav --stride 2.5

By default, detections are printed to the terminal in JSON format. To save detections as an Audacity label file use:

sauim-detector records/Mindu_Saguinus_bicolor_02.02.19-000.wav --save-detections

Save the band-pass filtered audio used by the detector:

sauim-detector records/Mindu_Saguinus_bicolor_02.02.19-000.wav --save-audio

Use all three options together:

sauim-detector records/Mindu_Saguinus_bicolor_02.02.19-000.wav --stride 2.5 --save-detections --save-audio

Outputs

With --save-detections, the CLI writes an Audacity-compatible label file named <input>_detections.txt next to the input audio:

start_time    end_time    label
0.00          7.20        Pied tamarin
10.00         15.50       Pied tamarin
20.00         30.80       Pied tamarin

These labels can be imported into Audacity with File > Import > Labels....

Without --save-detections, the CLI prints JSON detections to stdout. Each detection includes:

• species: common name
• scientific: scientific name
• confidence: One-Class SVM decision score, not a calibrated probability
• start_time: detection start time in seconds
• end_time: detection end time in seconds

Example:

[
  {
    "species": "Pied tamarin",
    "scientific": "Saguinus bicolor",
    "confidence": 0.003421,
    "start_time": 12.0,
    "end_time": 17.0
  }
]

Notes

• Input files must be .wav files.
• Audio is loaded at 32 kHz.
• The classifier only detects the target species, Pied tamarin (Saguinus bicolor).
• The confidence field is the raw OCSVM decision score. Positive scores are detections; negative scores are rejected before output.
• Labels should be manually validated in Audacity.