Attempt 3 listening: what the model learned on sessions it trained on
The same six 10-second excerpts as before. This time the model that trained on these sessions (in-sample) is included next to the held-out model. Press play, then switch versions while it plays. The switch is instant and keeps your place. On a computer, keys 1-9, 0, then Q W E R T Y and so on follow the buttons in order, and the space bar plays or pauses the card you last touched.
A is the raw audio and C is your final. Held-out B is the model that never saw the session (what you heard before). In-sample B is the model trained on all 11 sessions, so it has seen these exact excerpts. Memorized B (molly and duolingo only) is a copy trained for about half an hour on just those two sessions, with no regularization, to show the most it can copy. Memorized is not a real result, since it only works on audio it has already seen. The second row holds the deltas, meaning each version minus the raw audio, so you hear only what it changed. Every delta in a card uses the same boost, shown in the card, so loudness compares fairly. D is your change, and F is what the in-sample 1x model missed (B minus C). Audio is lossless 16-bit FLAC.