Outputs¶

The default segmentation model returns sigmoid scores in [0, 1]: channel 0 describes coherent activity and channel 1 describes transient activity. A score is not a binary mask; apply a threshold appropriate for the task.

TokEye.segment() returns a Segmentation with a (C, H, W) mask, spectrogram, axes, and metadata. Plot with the lowest frequencies at the bottom (origin="lower"). With a sample rate, frequency and time are measured in hertz and seconds; otherwise they are bin and frame indices.

The CLI writes a NumPy mask, a preview, and a JSON provenance file per input. --format npz writes a bundle containing the mask, spectrogram, and axes instead of the bare mask. Reload a bundle using Segmentation.load().