Data leakage
Data quality
Information entering training or model selection that should be unavailable for the evaluation being performed.
Leakage can produce an overly optimistic result. It can come from overlapping examples, preprocessing fitted on test data or repeated tuning against a supposedly final test set. Similarity and shared collection context matter as well as exact duplicates.
In practice
A model is tested on clips from the same recording it was trained on, making the reported result a weak measure of performance on new recordings.