Training vs Inference PCEN

Training and inference intentionally differ in clip-boundary policy.

Training

Training consumes pre-segmented examples. Boundary handling is conservative to avoid cross-chunk leakage:

• clips are not allowed to cross chunk boundaries
• edge clips may be dropped
• this preserves split integrity and avoids hidden context bleed

Inference

Inference treats each recording as continuous. Internal segmentation is a memory strategy, not a semantic boundary:

• clip extraction targets full temporal coverage
• segment boundaries are bridged logically
• overlap (3 s clips, 1.5 s hop) reduces boundary misses

Why It Matters

Expected and Intentional Difference

The distinction between training and inference clip handling is expected and desirable:

• Conservative training supports a robust evaluation setup with clean split integrity.
• Dense inference coverage improves recall on long soundscapes by bridging segment boundaries.

Implementation references:

• Training side: BirdBox-Train preprocessing
• Inference side: src/inference/utils/pcen_inference.py