TTS-AGI/moss-character-reference-voices

MOSS character reference voices (1336 voices)

1336 distinct synthetic character voices, each mined from a cluster of generated MOSS-VA-v2 character audio and auto-annotated by Gemini-3-Flash. For every cluster the model was shown the 3 cluster samples

• their automatic voice scores, chose the single most representative sample, and wrote a full casting-style profile.

Contents

• dataset.jsonl — one row per voice: cid, name, tagline, description, age, gender, register, timbre_profile (metallic/throat_guttural/falsetto/chest_voice/roughness/brightness, 0–5), distinctive_features, emotional_range, casting (classic_fantasy / sci_fi / mystery_horror / contemporary → role + delivery direction), tags, chosen_scores (AGEV/BKGN/genu/blend/voice/dur), chosen_caption, search_text, audio.
• audio/<cid>.mp3 — the chosen best-of-three demo clip for each voice.
• emb.npy — GTE-large-en-v1.5 embeddings of search_text (L2-normalized), bm25.pkl — BM25 index, meta.json — server metadata.
• server.py — FastAPI search server (vector similarity + BM25); index.html — demo search page.

Search

Free-text description → vector similarity (semantic, sentence-embedding over the concatenated timbre/features/emotional-range/genre/tags text) or BM25 (keyword). Top-N voices with their demo audio.

pip install fastapi uvicorn sentence-transformers rank_bm25
DS_DIR=. uvicorn server:app --host 0.0.0.0 --port 8778   # then open index.html

🔬 Research use. All voices, audio and data are fully synthetic (AI-generated; no real people). Annotations are automatic (Gemini-3-Flash) from audio + scores.