avoc quality starter
Starter weights for the "low quality" hint in anime-voice-collector (avoc): the chance that an anime voice clip is too noisy for a clean dataset (loud background, two people at once).
The app downloads this file once. It trains on the user's own labels starting from these weights and holds the weights back toward them, so the hint works from the first label.
What is in the file
quality-starter-v1.npz holds a logistic regression, nothing else (no audio, no clip features):
mean,scale(6144 each): how the features are scaled before the regression.weights(6144),offset: the regression.layers:[4, 6, 8].
The features are, for each clip of 16 kHz separated vocals, the mean and the spread over time of each number in torchaudio's WAVLM_LARGE transformer layers 4, 6 and 8 (input scaled to mean 0 and spread 1).
Training data
Labels made by one person on two anime: 650 clips of Tokai Teio (Uma Musume) and 274 clips from usagi, each marked clean or low quality. Class weights balanced.
How good
Scored on usagi without any usagi labels, a model trained on Teio alone ranked a random low clip above a random clean one 91% of the time (auc 0.909).