crate-clap-general — CLAP for Core AI (macOS 27)

LAION's larger_clap_general (Wu et al., "Large-scale Contrastive Language-Audio Pretraining", ICASSP 2023; Apache 2.0) exported to an Apple Core AI .aimodel for on-device sample-library search, similarity and zero-shot tagging. Why this checkpoint and not larger_clap_music: the music checkpoint's Hugging Face conversion collapses every clip to nearly one vector (audio–audio cosines 0.81–0.99, audio–text 0.01–0.04, logit scale 1.03, measured in transformers itself), so it cannot search or tag anything. The general checkpoint was trained on music too and behaves (a kick file scores 0.49 against "a kick drum").

Read by swift-sample-search and the crate command-line tool.

Files

Path	What
crate-clap-general-float32.aimodel/	Two entry points: audio (log-mel [1,1,1001,64] → [1,512], HTSAT + projection) and text (RoBERTa ids [1,77] + mask [1,77] → [1,512]). Float32, 797 MB.
clap-support/vocab.json, merges.txt	The checkpoint's own RoBERTa byte-level BPE files.
clap-support/scales.json	exp(logit_scale_a) and exp(logit_scale_t) from the checkpoint (38.66 and 14.29), the token length, the source model id.

The host computes the log-mel exactly as ClapFeatureExtractor does (48 kHz, 64 Slaney mels 50–14000 Hz, n_fft 1024, hop 480, repeatpad for clips under ten seconds), tokenises with the files above, and L2-normalises the outputs — get_audio_features / get_text_features.

Fidelity

Measured against transformers 5.17 on the same inputs: tokenizer ids identical over 20 phrases; log-mel 116–156 dB PSNR; text embeddings 142 dB; audio embeddings 144–147 dB (cosine 0.99999+); zero-shot probabilities within 1e-8. The export script and the fixtures are in the library repo (Tools/export_clap.py).

Use

hf download arraypress/crate-clap-general --local-dir models
crate model install models
crate index ~/Samples && crate search "punchy 808 kick"

Requires macOS 27 (Core AI) on Apple silicon. Converted with coreai-torch 0.4.2 / torch 2.13.