laion/moss-local-full-reinterpretations-64

MOSS-Local Full-Performance Reinterpretations — best-of-64, dual-reward scored

2,953 full two-scene voice-acting performances (EN/DE/ES/FR, all 9 prompt pathways), each reinterpreted 64 times by laion/moss-tts-local-transformer-4.55b-voice-acting (4.55B local transformer, native 48 kHz), voice-cloned from the original performance, with the complete scoring needed to select the best takes two different ways — voice identity or emotional performance.

Source performances: TTS-AGI/dramabox-gemini-finetune (Gemini-prompted two-scene performances, "CUT TO:" format). Methodology mirrors the reproduce-and-improve study, scaled from best-of-8 to best-of-64.

How each group was made

• instruction = the sample's full Gemini performance prompt (voice description + stage directions + quoted dialogue with vocal-burst notes, both scenes incl. "CUT TO:")
• text = scene-1 + scene-2 expected texts
• reference = the ORIGINAL full performance audio (both parts, in order) — encoded to 12-codebook MOSS v2 codes for voice cloning, and kept as the comparison target
• generation: audio_temperature 1.0 (with-reference setting), top-p 0.95, top-k 25, repetition 1.1, batched 64-at-once, token budget = max(words×6, ref_frames×1.2) at 12.5 Hz
• audio: raw native 48 kHz FLAC — no post-processing

Per-take scores (scores.parquet, also per-bucket inside each tar)

column	meaning
wer, inv_wer	word error rate vs the expected text (nvidia/parakeet-tdt-0.6b-v3, bf16 batched)
genu, blend	genuineness & vocal-burst-blend heads on laion/voiceclap-commercial
prompt_sim	VoiceCLAP cosine of direction-text vs audio
ei_* (42 cols)	Empathic-Insight-Voice-Plus scores: 40 EmoNet emotions + Valence + Arousal (laion/Empathic-Insight-Voice-Plus on laion/BUD-E-Whisper)
emonet_42	the same 42 values as one vector (list)
emotion_cos	cosine(take's 42-vec, reference's 42-vec) — "same feeling?"
ecapa_192	ECAPA-TDNN speaker embedding (192-d list, speechbrain/spkrec-ecapa-voxceleb)
spk_sim	cosine(take's ECAPA, reference's ECAPA) — "same voice?"
dur_match	min(dur, ref_dur)/max(dur, ref_dur)
rms_db, peak_db, dur	loudness & length

references.parquet holds each group's reference 42-dim EmoNet vector + ECAPA embedding.

The two rewards (compute from the raw columns; min-max normalise globally)

reward_A (voice identity)     = mean( n(spk_sim), n(inv_wer), n(genu), n(blend), n(dur_match) )
reward_B (emotion profile)    = emotion_cos × inv_wer

Ranking A asks "same speaker, intelligible, at least as human as the original?"; Ranking B asks "does it feel like the same performance?". They disagree often — that is the point.

Layout

data/bucket_XXX.tar      ~100 groups each: <gid>/<gid>_vNN.flac (64 takes, 48 kHz)
                         + bucket_XXX_scores.parquet
scores.parquet           all takes, all scores + embeddings
references.parquet       per-group reference vectors/embeddings

Generation stack

Model laion/moss-tts-local-transformer-4.55b-voice-acting · codec OpenMOSS-Team/MOSS-Audio-Tokenizer-v2 · recipe & code: github.com/LAION-AI/laion-moss-local-1.5-voice-acting-4.55b · 8×A100, fused generate→score workers (~1.3–1.6 s/clip/GPU incl. all scoring).

License

CC BY 4.0. Synthetic speech generated by the model named above; source prompts/performances from TTS-AGI/dramabox-gemini-finetune (CC BY 4.0).

---

The source performances were generated with a voice-acting AI model.