TTS-AGI/acting-transitions-200

Acting transitions — is this join human, or is it an edit?

200 part-to-part transitions from synthetic acting performances, each rated 0–5 by gemini-3.6-flash for how naturally the two pieces join, with VoiceCLAP embeddings and discontinuity features attached.

Every sample is three audio files — part A alone, part B alone, and the two joined — so the dataset supports the thing a single concatenated clip cannot: telling "these two performances are different" (intended) apart from "the join is bad" (the fault).

Why this exists

A text-to-speech model cannot produce 30 seconds of dramatic acting in one generation. It truncates, and a single generation cannot carry two opposed emotions. So a long performance is built from 2–4 separately generated parts and joined with a crossfade.

That creates a seam. The script wants the emotion to change across it — that is the drama — but it must not sound like an edit: a different person, a different room, a jump in loudness or pace. This dataset measures whether it does, and provides features to test whether the failure is detectable without listening.

The rubric (verbatim, as given to the judge)

The judge received three audio clips per request, in this order: part A, part B, then A+B joined. It was told that a large emotional change is not a fault.

5 = Indistinguishable from a single continuous take by a very good human actor. Even a large emotional change lands organically. Vocal bursts blend into the speech. Same person throughout. Breath, level, room and pace carry across the join. 4 = Natural. A listener would not question it. Perhaps a slight seam if actively listening for it. 3 = Acceptable but noticeable: a small discontinuity in energy, pace or emotional state. 2 = Clearly two takes edited together. The emotion jumps rather than moves. 1 = Jarring. Abrupt emotional switch, level or timbre shift; sounds like a different recording. 0 = Completely disjoint or chaotic; plausibly a different speaker or an unrelated performance.

Output format: SCORE, SAME_SPEAKER (yes/no), DISCONTINUITY (none/mild/severe), and a one-sentence justification — all retained per sample.

Judge: gemini-3.6-flash, a reasoning model, thinking enabled, temperature 0. Audio sent as 48 kbps mono MP3 via the native generateContent endpoint with inline_data. Clips were encoded at up to 90 s so no concatenation was truncated — the OpenAI-compatible chat endpoint silently drops audio and answers anyway, which is why the native endpoint is used, and this was re-verified before the run by asking the model to describe what it heard.

Generation conditions

Transitions were generated fresh rather than harvested, because no per-part audio existed — the source pipeline only ever saved finished, already-joined takes, and seam positions inside them are not recoverable. Generating fresh also allowed the quality range to be designed rather than hoped for: a judged set containing only good transitions cannot demonstrate that a feature detects bad ones.

Five conditions, assigned round-robin over the source plans:

arm	what part B is conditioned on
cont_anchor	audio continuation (anchor + last 4 s of A in the assistant turn) plus an explicit "the same speaker continues … no cut, no new narrator" sentence appended to the caption. The production path.
cont_notext	identical audio conditioning, continuity sentence removed — isolates the contribution of the words
cont_hot	continuation, but B's emotion adapter forced to 0.9 and the caption demands the voice snaps hard out of the previous emotion
reference	part A supplied only as reference= audio, no continuation — the historical "cut between two recordings" path
independent	no conditioning at all — an unrelated draw of an unspecified speaker. The deliberate negatives.

Source material: archived performance plans from nine autonomous casting agents (real GENERAL/SCRIPT captions, emotion and vocal-burst adapters, emotional targets, declared prosody turns). Base model laion/moss-tts-local-transformer-4.55b-voice-acting-v2 with emotion and vocal-burst adapters.

The join is made exactly as the production assembler makes it: part B level-matched to part A over the loudest 60 % of frames, then a 25 ms equal-power crossfade, then one gentle peak normalisation.

Stratification

The 200 were selected from ~400 generated pairs, stratified two ways:

1. Task — the nine acting challenges, so no single scene dominates and a model can be evaluated with a held-out-scene split.
2. Quality — five quantile buckets of a cheap proxy, so the annotated set spans the range before the judge sees it.

The proxy is 2·z(ECAPA speaker similarity) − 0.7·z(|Δtempo|) − 0.7·z(|Δchunking|) − 0.4·z(|Δdisfluency|) − 0.6·z(|Δlevel dB|).

The proxy deliberately excludes every VoiceCLAP feature under test. Selecting samples with the same features later claimed to be predictive would be circular.

Contents

path	what
data/transitions-0000.tar	WebDataset: <uid>.partA.flac, <uid>.partB.flac, <uid>.concat.flac, <uid>.json
data/transitions-0000.parquet	flat manifest (all scalar fields)
manifest.parquet	same, top level
clap_embeddings.npz	uid → (9, 768) float32

Per-sample fields

Label — score (0–5), same_speaker, discontinuity, why, gemini_raw.

Provenance — uid, task, run, rnd, pi, arm, scene, textA/textB, captionA/captionB, adaptersA/adaptersB, targetsA/targetsB, turn_from, tempo_target*, chunk_target*, prosody_turn, durA, durB, durCat, seam_s.

VoiceCLAP embeddings — clap_embeddings, shape (9, 768), in order:

index	content
0	w0 — 10 s of A ending at the seam (pure A)
1	w1 — 8 s A + 2 s B
2	w2 — 6 s A + 4 s B
3	w3 — 4 s A + 6 s B
4	w4 — 2 s A + 8 s B
5	w5 — 10 s of B from the seam (pure B)
6	e_A — whole part A
7	e_B — whole part B
8	e_boundary — last 5 s of A + first 5 s of B

Encoder: VoiceCLAP commercial, the same 768-d encoder the project's VoiceNet, genuineness and vocal-burst-blend heads sit on. Windows 0–5 are all exactly 10 s, so they are directly comparable; the encoder zero-pads to 30 s either way. Indices 6 and 7 vary in duration — treat whole-part cosines as carrying a length confound that the fixed-length trajectory does not.

Derived features — cos_AB, cos_w0_w5, cos_bound_A, cos_bound_B, bound_off_mid, step0…step4 (1 − cosine between consecutive windows), step_max, step_mean, step_std, step_range, step_concentration (largest step ÷ total path length; 1.0 = all movement in one step, i.e. a cut), curvature, traj_total.

Independent sensors — spk_sim (ECAPA), tempoA/B, chunkA/B, dfluA/B (VoiceNet TEMP/CHNK/DFLU), genuA/B, blendA/B, level_db, and the d_* deltas.

The hypothesis being tested

A performance that genuinely continues should trace a gradual path through embedding space from "sounds like A" to "sounds like B". A hard cut should show one large step concentrated at the boundary.

step_concentration is the direct operationalisation of this.

Results

See RESULTS.md in this repository and the interactive report (every sample with audio players, the score distribution, trajectory plots and the model table).

All model numbers are out-of-fold under GroupKFold by task — the model never sees the scene it is tested on. A random split would let it memorise a scene's voice and report a far better number for no real reason. With n=200 and 768-d embeddings a neural network fits the training set perfectly and means nothing, so the reported comparison is against the single best hand-made feature, with bootstrap confidence intervals.

Caveats

• All audio is synthetic, including any apparent room or recording character.
• Labels are model judgements, not human ratings. No human listening study was run. The judge is one model with one prompt; systematic biases in it become systematic biases here.
• The five generation conditions make the score distribution deliberately non-natural — it over-represents bad joins relative to what the production pipeline actually emits. Do not read the distribution as a quality estimate of the pipeline.
• n=200 is small for 768-d representation learning. It is sized to answer "is there signal", not to train a deployable detector.