clap-htsat-base-onnx

muzaiten/clap-htsat-base-onnx

clap-htsat-base-onnx is a machine learning model from muzaiten. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for onnx. The card lists the license as cc0-1.0.

Pre-converted ONNX Runtime artifacts of the LAION-CLAP music checkpoint musicaudiosetepoch15esc90.14.pt (HTSAT-base audio tower, RoBERTa text tower, non-fusion), packaged for muzaiten's optional semantic-analysis prov…

Downloads · 30 days

Access

Public

Updated Jul 12, 2026

Repo size

787 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.onnx787 MB · 100%

At a glance

Library: onnx
License: cc0-1.0
Access: Public
Created: Jul 12, 2026
Updated: Jul 12, 2026
SHA: d4d4372c

Model card

From the Hugging Face model README

CLAP HTSAT-base music model, fp32 ONNX bundle

Pre-converted ONNX Runtime artifacts of the LAION-CLAP music checkpoint music_audioset_epoch_15_esc_90.14.pt (HTSAT-base audio tower, RoBERTa text tower, non-fusion), packaged for muzaiten's optional semantic-analysis provider muzaiten-features-clap.

The upstream code and checkpoint are CC0-1.0; this bundle is a mechanical conversion and carries the same license.

Files

File	Contents
`audio.onnx`	HTSAT-base audio tower plus projection; input `waveform` float32 `[batch, 480000]` (10 s at 48 kHz, mono), output L2-normalized `[batch, 512]`; mel spectrogram computed in-graph
`text.onnx`	RoBERTa-base text tower plus projection; inputs `input_ids`/`attention_mask` int64 `[batch, 77]`, output L2-normalized `[batch, 512]`
`tokenizer.json`	The exact RoBERTa tokenizer used by the checkpoint
`manifest.json`	SHA-256 and byte count of each artifact plus the source checkpoint hash (`fae3e9c0…`) and feature revision (`clap-htsat-base-audio-window-v1`)

Provenance and parity

Exported with muzaiten's converter (python -m muzaiten_features_clap.convert, torch legacy exporter, opset 18) from the SHA-256-pinned upstream checkpoint. Full-pipeline parity against the reference laion_clap implementation: 68 audio cases with mean cosine 1.0 and minimum 0.9999989; text token ids and attention masks exactly equal with cosine 1.0. Embeddings are interchangeable with the torch reference, so muzaiten feature stores need no re-analysis.

muzaiten-features-clap downloads this bundle through model download, rejects the manifest unless its identity matches the constants baked into the provider release, and verifies every file hash while streaming. See muzaiten's semantic-analysis guide for usage; direct consumers can load the graphs with plain onnxruntime and tokenizers, replicating the int16 round-trip quantization and repeat-pad preprocessing described in the provider source.

Library: onnx
License: cc0-1.0
Created: Jul 12, 2026
Updated: Jul 12, 2026