README.md

metadata

license: cc-by-4.0
task_categories:
  - text-to-speech
  - audio-classification
language:
  - en
  - de
tags:
  - speech
  - voice-acting
  - tts
  - best-of-n
  - reward-ranking
size_categories:
  - 100K<n<1M

MOSS Character Voices — Best-of-64 (Stage 2)

Best-of-64 voice-acting takes from the 4.55B MOSS-TTS-Local voice-acting model (laion/moss-tts-local-transformer-4.55b-voice-acting) for 13 evolved character voices.

Each prompt is a fixed, optimized champion performance direction (instruction) paired with a Gemma-generated topic text (text) — together, one performance to render. For every prompt we sample 64 takes with distinct seeds at 48 kHz, score each take, and rank the 64 within the group (rank 1 = best) by a per-character reward.

Scoring

Each take is scored with the production no-reference stack:

• VoiceCLAP-commercial — blend (0–10 naturalness/quality) and genu (genuineness 0–6).
• 57 VoiceNet regression heads — raw acoustic/signature dims (dims_json); each character weights a subset of positive/negative dims into char.
• Per-character EmoNet emotion heads (BUD-E-Whisper encoder + Empathic-Insight-Voice-Plus) — raw emotion scores (emo_json) weighted into char_emo.
• Parakeet-TDT ASR → invwer (1 − WER, clamped).

The per-character reward = w_blend·norm(blend) + w_genu·norm(genu) + w_vn·char + w_emo·char_emo, where every term is min-max normalized WITHIN the group of 64, then multiplied by a WER factor (gate/soft/none, per character). Takes are ranked by reward (rank 1 = best).

Layout (WebDataset)

• webdataset/shard-NNNNN.tar — FLAC 48 kHz (lossless, PCM_16), keyed {group_uid}_{rank:02d}.flac.
• metadata/scores-NNNNN.parquet — one row per take (paired with the same shard): group_uid, character, lang, topic, text, instruction, flac_key, seed, rank, reward, char, char_emo, blend, genu, invwer, dur, hyp, dims_json, emo_json, best_reward, mean_reward, best_key.

Progress (live)

12,901 / 12,984 prompt-groups complete · 825,274 takes · 130 shards.

character	groups
asmr_man	1000
asmr_woman	1000
dragon	996
evil_ghost	999
fairy	993
goblin	998
mouse	1000
ork	998
pain_scream	1000
ranting	1000
sad_man	1000
sad_woman	1000
zombie	917

lang	groups
de	6464
en	6437

Scale

Target: 12,984 groups × 64 = ~831k clips, ~250–380 GB FLAC. Generated on 8× A100/H100 over ~2–3 days; uploaded incrementally in ~1 GB shards (dataset grows as generation proceeds).

Characters

asmr_man, asmr_woman, dragon, evil_ghost, fairy, goblin, mouse, ork, pain_scream, ranting, sad_man, sad_woman, zombie — each a distinct evolved voice with its own reward configuration.

License: CC-BY-4.0.