FastAPI server + web demo to search 6,064 reference voices (MOSS character voices + EmoLia real/multilingual
- Japanese-anime-derived + reinterpreted + German-Mediathek-derived) by free-text description and filter on any of 99 measured dimensions (57 VoiceNet voice-quality + 40 Empathic-Insight emotions + genuineness + burst-blend), returning the top-N.
- Demo:
docs/index.html→ https://projects.laion.ai/moss-reference-voice-search/ (query + top-N selector + multi-condition dimension filters). - Server:
server/server.py(FastAPI). Run:CUDA_VISIBLE_DEVICES=0 DS_DIR=<index> uvicorn server:app --host 0.0.0.0 --port 8778. - API: see
API.md—/search(q, mode, n, filters, with_dims),/dims(dimension catalog),/audio/{cid}.mp3. - Dimensions:
dim_catalog.json— every filterable dimension with code, name, group and a one-line description.
The index data (meta.json, emb.npy, vclap_audio.npy, dims.npy, bm25.pkl, audio/) is served alongside the running server; the demo talks to the server over a Cloudflare tunnel. 🔬 Research; fully synthetic/auto-annotated.