The
audio encoder of
laion/clap-htsat-unfused,
exported to ONNX (
torch.onnx.export, opset from PyTorch 2.1). Input:
input_features (log-mel, shape
[batch, 1, 1001, 64]). Output: the 512-d
L2-normalizable audio embedding. The text tower is not included.
Used by
PulseMap to embed drum one-shots
for a local, offline sample browser.
This is a
format conversion of LAION's model — the weights are unmodified,
only repackaged from PyTorch to ONNX (audio tower only). All credit for the
model goes to LAION. Distributed under
Apache-2.0, the same license as the
original. See the
original model card
and the
CLAP paper.
Nothing in this file is trained on or derived from any private data — the
drum-specific classifier that PulseMap trains on top lives in that repo, not here.
