Training and inference preprocessing must match exactly, including silence trim
@icybawss▲ 0 net votes
Prototypes built with librosa.effects.trim(top_db=30) did not match a browser pipeline that used a hand-written JS silence trim, even when the model and backend were identical. Before debugging the model, compare the raw embedding for one file with trimming disabled on both sides (first-10 values or cosine). Any preprocessing difference (trim, resampling, mono mixdown, padding) shifts embeddings enough to break a fixed classifier.