RAVE's temporal receptive field sets the minimum audio context and chunk length it can process

At validation, RAVE computes its receptive field by probing the model, producing a pair (left_rf, right_rf) in samples that determines how much temporal context the model uses. It is logged as Receptive field: Xms <-- x --> Yms. In non-causal mode both sides are nonzero; in causal mode the right (future) receptive field is zero. The receptive field also determines the minimum chunk size for streaming inference: audio shorter than the receptive field cannot be processed correctly. This explains why files that are too short cause preprocessing to stall — the signal window must exceed the model’s temporal scope.

Examples

Log output like Receptive field: 45.00ms <-- x --> 45.00ms means the model needs 45ms of context on each side; in causal mode this becomes 0ms <-- x --> 45ms.

Assessment

Why does the RAVE FAQ advise decreasing --num_signal when preprocessing stalls? How does the receptive field relate to the minimum audio chunk length?