CLAP (Contrastive Language-Audio Pretraining)

Multiple (Microsoft/LAION-style CLAP architecture)

Structures found in this family (3)

Linear Subspace, Anisotropy, Linear Direction

By model (3)

CLAP (HTSAT-BERT-ZS, trained on WavCaps)

Linear Subspace, Anisotropy

DRCap's CLAP (trained on WavCaps + SoundVECaps)

Linear Subspace, Anisotropy

LAION-CLAP (HTSAT backbone, trained on LAION-Audio-630k + AudioSet + speech/music)

Linear Direction

Observations (2)

Audio-text embeddings split into shared head and private tail

CLAP linearly encodes reverberation and loudness with cross-dataset-consistent axes

Papers

COMET: Concept Space Dissection of the Modality Gap in Audio-Text Multimodal Contrastive Embeddings (2026), Probing Low-Level Acoustic Attribute Encoding in CLAP Audio Embeddings (2026)