Contrastive Language-Audio Pretraining (CLAP) model that recognizes similarities between audio-text pairs, ingesting Orcasound community reports and Orcahello moderator annotations as training data. Primary applications include natural language search of audio files, and annotation assistance.

Proposed layout:

• python/ (inference server—tiny FastAPI with /embed, /score, /nearest)
• tools/
  • node-text-audio-pairs/ (Node utility for generating pairs)
  • shared CSV schema samples
• models/ (download/readme scripts, no weights checked in)
• docs/ (usage + examples)

APIs this repo should expose:

• POST /score → { audio_url | wav:bytes, prompts: [text...] } -> { scores: [{prompt, sim}], version }
• POST /nearest → { audio_url, k } -> { neighbors: [{audio_id, sim, meta}], version }
• POST /caption (optional/MVP later) → { audio_url } -> { caption, evidence_neighbors }