laion/dramabox-reinterpretations-top3

DramaBox reinterpretations — reward-ranked top 3 of 64

~20,000 acting prompts, each re-performed 64 times by laion/moss-tts-local-transformer-4.55b-voice-acting-v2, with the three highest-reward takes published here — plus the original DramaBox clip each one reinterprets. Roughly 80,000 audio files (~20k originals + ~60k takes).

The companion release …-raw64 keeps all 64 candidates per group, so the full reward distribution — not just its upper tail — stays available.

⚠️ Read this first: the texts are partly gibberish

The scripts come from a synthetic acting corpus and are a mix of English and German with many disfluencies, mid-word breaks and non-sequiturs. A typical line:

"Morgan, we are in a negotiation of physical gravity and climbing, which, while just div- any negotiation…"

This is a property of the source material, faithfully reproduced by the reinterpretation. Practical consequences:

• Good for: ASR training (the transcripts are accurate against difficult audio), prosody and expressive-delivery modelling, vocal-burst detection, best-of-N reward research.
• Bad for: anything that assumes the text is well-formed language.

Median WER against the intended script is 0.17, and only 1.6 % of candidates transcribe exactly. That is largely the disfluent source text, not model failure — the acoustic quality is much better than the WER suggests.

Scope

Complete: ~20,000 groups over both halves.

• the edge-case half — 10,000 groups over 262 distinct edge-case categories (ec000 … ec261), roughly evenly;
• the prod1M half — ~10,000 groups over the non-edge DramaBox material (the Gemma-authored acting prompts, stratified by pathway).

Each group now ships the original DramaBox clip alongside its three reinterpretations, so a take can be compared against what it reinterprets without resolving src_id anywhere else.

Cite the number in manifest.parquet rather than a round figure.

Contents

path	what
data/top3-NNNN.tar	WebDataset shards: one .mp3 + one .json per clip
data/top3-NNNN.parquet	per-shard manifest (key, reward, WER, duration, blend, genuineness, burst count)
manifest.parquet	all shards concatenated

Key format: <src_id>.top<rank>, rank ∈ {1,2,3} by descending reward, plus <src_id>.orig — the original DramaBox source clip for that group, so every group is four audio files. The original's JSON carries is_original: true, rank: 0 and the prompt, but no reward or model scores (it is not a candidate). Filter it out with is_original or rank > 0 when you want only the generations.

Per-clip JSON

reward_v1c, reward_parts (s/t/b/g), rank_in_group (position among all 64), wer, duration, asr_text, words (word-level timestamps), caption, blend_0_10, genuineness_0_6, emo_sim, emonet (40 dims), voicenet (57 dims with regression value, bucket and natural-language label), quality (4 dims), bursts (spans with start/end/duration and type), plus the full prompt (prompt_caption, prompt_general, prompt_script) and the source identity (src_id, src_duration, category).

The original source audio IS included as <src_id>.orig.mp3 (see key format above), so no lookup against the DramaBox staging corpus is needed.

The LoRAs used

The reinterpretations were generated with the v3 emotion LoRA adapters, rank 32 / alpha 64, trained against the same v2 base:

TTS-AGI/moss-emotion-loras-v3 — 40 emotions, one adapter directory each. Eight of them (Anger, Fatigue_Exhaustion, Fear, Malevolence_Malice, Pain, Sadness, Sexual_Lust, Teasing) drive the edge-case half of this corpus, hot-swapped per group; adapter swap costs ~0.021 s over 268 modules, so switching between groups is free relative to generation.

Base model: laion/moss-tts-local-transformer-4.55b-voice-acting-v2. These adapters are trained against v2 and will not behave correctly on the earlier checkpoint.

🎧 Listen: 25 originals vs their top-3 reinterpretations

The reward

R = (1.00·s + 1.25·t + 1.00·b + 1.00·g) / 4.25 · (1 − min(WER, 1))

s = (cos + 1)/2 over a 42-dim profile (40 EmoNet dims + arousal + valence),
    measured against the SOURCE clip — "does it re-act the same emotion"
t = sigmoid(target emotion, z-scored WITHIN the 64-candidate group)
b = sigmoid(vocal-burst blend, corpus robust-z)
g = sigmoid(genuineness, corpus robust-z)

Two design choices worth knowing, because both were arrived at by measurement:

1. The WER gate is multiplicative, never /(1 + WER). About three quarters of candidates have a negative core score, and dividing a negative number by a larger denominator makes it larger — so the division form literally rewards transcription errors on most of the pool. This was confirmed independently on a separate 1,000-clip set, where a literal WER × quality filter put 602 of 1,000 candidates at exactly score 0 (60 % had WER 0.00) and therefore selected the half with worse transcription. Squashing each component through a sigmoid first makes the core strictly positive, so × (1 − WER) is monotone decreasing in WER for every candidate.

2. The target emotion is z-scored within each group, not against the corpus. 20 of the 40 EmoNet dimensions have a corpus MAD ≤ 0.02 and five are exactly 0.0 — the corpus is overwhelmingly not any given emotion, so a corpus robust-z divides by ≈0 and explodes. Within group is also the right semantics: the task is ranking inside a group. Blend and genuineness are well-conditioned (MAD 0.625 and 1.601 over 256k clips) and keep corpus robust-z so they stay comparable across groups.

Observed distributions

Over a 38,400-candidate sample of the full pool (i.e. before top-3 selection):

	mean	p10	median	p90
reward_v1c	0.411	0.265	0.424	0.560
WER	0.329	0.067	0.170	0.444
blend_0_10	2.154	0.000	1.656	4.966
genuineness_0_6	1.656	0.635	1.503	2.945
duration (s)	15.45	9.44	14.40	21.36

46.8 % of candidates contain at least one detected vocal burst (mean 0.81 per clip).

Annotation

• ASR: CrisperWhisper 2 large, verbatim mode, word-level timestamps. Chosen over faster models specifically because the transcripts are a deliverable here, not just a ranking signal.
• Vocal bursts: located first, then classified — with a 200 ms minimum duration floor. The span timestamps are the reliable part; the class names are a weak model prior. The classifier names an already-isolated burst and is known to be unreliable on bursts embedded in continuous speech, where it drifts toward sigh/gasp/ahem. Filter on span timing, treat the label as a hint.
• Emotion/voice attributes: MLP heads over a VoiceCLAP encoder (EmoNet 40, VoiceNet 57, quality 4, genuineness, blend). Model outputs, not human ratings.

Caveats

• Everything here is synthetic audio generated by a TTS model, including any apparent background or recording character.
• Scores are model outputs throughout; no human evaluation was performed on this corpus.
• Top-3 selection means this release is the upper tail of the reward distribution by construction. For unbiased analysis use the raw-64 companion.
• The top3 flag was verified against a fresh ranking by reward_v1c on 400 groups — 400/400 agreement. Note that this check sampled the edge half only, which is exactly why it did not catch the issue below.

A ranking bug that affected the prod1M half (fixed before release)

Worth stating plainly, because it is the kind of failure that ships silently.

The reward's similarity term s compares a candidate against the source clip's emotion profile, so it needs the source clip scored. The edge half carries those scores already; the prod1M half does not, and the step that computes them (src_scores_prod1m.parquet) was never run. The annotator's reward block is guarded by a lookup of that table, so for every prod1M group it was skipped, leaving each candidate without a reward_v1c. Ranking then fell through to sorted(..., key=lambda i: -c.get("reward_v1c", -1e9)) — every candidate tied, Python's sort is stable, and the input order survived as the "ranking". The "top 3" for that half was candidates 0, 1, 2 of 64.

Nothing downstream could tell: three candidates were flagged per group, rank_in_group was populated 0…63, and the schema validated. Only the values gave it away — a real ranking is a permutation, and this one was the identity.

Resolution. The source vectors were computed, reward_v1c was recomputed for every affected group, and the shards were rebuilt. The packer now refuses to run if the source-score table is missing, and a gate samples the rescored groups and aborts if top3 is still [0,1,2]. The published shards are ranked by the reward described above throughout.

If you downloaded this dataset before 2026-08-08, the prod1M shards you have are arbitrary picks rather than top-3. The edge-half shards were always correctly ranked.

License

other. Derived from synthetic generations over the DramaBox acting-prompt corpus. Check your own jurisdiction before use beyond research.