Datasetpublic

laion/acting-challenge-dataset

Acting Challenge Dataset 19,247 English acting challenges for constructing expressive speech tasks and synthetic data. This is a published inspiration and training resource, not the Voice Acting Arena evaluation set. A challenge specifies a dramatic or physical performance problem, usually before any exact spoken script has been written. A downstream director model can turn it into a scene, dialogue, and model-specific TTS directions; a generator can produce several takes; human… See the full description on the dataset page: https://huggingface.co/datasets/laion/acting-challenge-dataset.

sourceHugging Faceapache-2.0updated 14h agoView on Hugging Face

0likes

Dataset Card

Acting Challenge Dataset

19,247 English acting challenges for constructing expressive speech tasks and synthetic data. This is a published inspiration and training resource, not the Voice Acting Arena evaluation set. A challenge specifies a dramatic or physical performance problem, usually before any exact spoken script has been written. A downstream director model can turn it into a scene, dialogue, and model-specific TTS directions; a generator can produce several takes; human or audio-model feedback can select takes for best-of-N sampling or later training. The same scene concepts may also be useful for prompting other generative media, but no video-generation evaluation is claimed here.

This repository republishes a pinned snapshot of the challenge pool from Voice-Acting-Pipeline, with a viewer-friendly JSONL file, source-family sampling weights, a small deterministic family-balanced reference view, and the source taxonomies. The source snapshot is commit e3b0d5281558f7f73e1118c6a9b4c149bb8c0621; the source master JSON has SHA-256 7ef759b639ef3646ca8a985664747e37ad93e4ab989f97dd8cece4f5201ceb6f.

This dataset is indexed in the Humaneness Voice: Research Artifacts and Voice Acting: Training and Preference Views collections.

What is in the release

Source family (`source`)	Rows	Construction and role
`laions_got_talent`	278	The original LAION's Got Talent acting tasks, with varied emotional and situational contexts. This is not the television programme Britain's Got Talent.
`haiku_batch`	1,000	AI-generated challenges intended to broaden thematic coverage. The source documentation does not identify a reliable historical generator model.
`voicenet_extension`	200	Scenarios built around VoiceNet vocal-attribute dimensions.
`eric_morris_inspired`	4,030	Five generated challenges per item in an 806-item historical snapshot of concepts, techniques, exercises, and recommendations from an Eric Morris acting-method knowledge base. `inspired_by` names the source concept; the seven books and their extracted text are not redistributed here.
`existing_challenge_inspired`	7,390	Five contextual variants for each of the 1,478 original challenges (278 + 1,000 + 200). These are deliberate descendants, not independent human-authored seeds.
`situation_inspired`	5,749	Up to 20 variants per each of 289 situations in 11 physical, social, and acoustic dimensions; the theoretical 5,780 were reduced by 31 failed generations. Each variant sampled one to three EmoNet emotions and intensity, plus age and gender. The source pipeline documents DeepSeek V4 Flash for this step.
`extreme_physical`	600	Six physical-condition categories, ten subcategories each, ten scenarios each: exertion/tension, breathlessness, pain, temperature, oral/taste sensation, and startle. These are designed to create large performance contrasts, not a representative distribution of real-life speech.
Total	19,247	Seven source families, mixed and shuffled in the upstream master.

The currently visible Morris knowledge-base index has 799 items, whereas the challenge-generation snapshot is documented as 806. The source does not explain the seven-item difference. The arithmetic 806 × 5 = 4,030 verifies the historical generation count, not the present-day index. Generator identities for the historical Morris, existing-challenge, and extreme-physical expansions are not reliably documented and are therefore not asserted. Gemma 4 E4B-it is documented for the later conversion of challenges into DramaBox-style scripts, which are not included in this dataset.

Balance, provenance, and configurations

The full configuration retains every original row. It is not family-balanced: the 7,390 legacy variants and 5,749 situation variants are much larger than the 200-item VoiceNet extension. Each row has family_sampling_weight = 1 / (7 × number_of_rows_in_its_source_family). Sampling proportionally to these weights makes the seven source families equally likely. This does not make emotions, speakers, language, or all taxonomy intersections equally frequent.

The family_balanced_reference configuration is a reproducible 1,400-row view with 200 examples from each family. It samples by stable SHA-256 order and round-robins over inspired_by seeds for Morris and legacy variants, situation dimensions for situation-derived tasks, and physical categories for extreme tasks. It is useful for quick inspection or small controlled prompt tests, not an independent evaluation split. Its rows are part of the full configuration. The selection procedure is in `build_dataset.py`; counts and source hash are in `MANIFEST.json`.

The source master is kept byte-for-byte in `source/all_acting_challenges.json`, together with the separately published Morris, legacy, situation, and extreme JSON files. Source fields such as id, title, instruction, source, inspired_by, situation_dim, situation_name, and emotions_sampled are preserved. The normalized JSONL enriches the 600 extreme tasks with their category_code, subcategory, scenario, and sampled speaker/emotion fields from the original extreme file. It adds only family_sampling_weight; no challenge text is rewritten. id identifies a row in this shuffled snapshot, not a permanent cross-version identifier. Many titles recur; title is not a key.

The accompanying `taxonomies/` directory contains the source EmoNet taxonomy (40 emotions), Situation Taxonomy (289 entries in 11 dimensions), Extreme Physical Taxonomy (6 × 10 × 10 construction), and the source VoiceNet extension HTML. The 57-dimensional VoiceNet taxonomy is described by the upstream documentation. acoustic_signature and related situation descriptions are prompt-design hypotheses, not validated medical or phonetic labels.

Example use

python

from datasets import load_dataset

repo = "laion/acting-challenge-dataset"
all_rows = load_dataset(repo, "full", split="train")
quick_view = load_dataset(repo, "family_balanced_reference", split="train")

challenge = all_rows[0]
director_input = f"Acting challenge: {challenge['title']}\n{challenge['instruction']}"

For a full-data sampler, use the family_sampling_weight field; for controlled studies, also stratify within the situation and emotion metadata rather than assuming the full pool is jointly balanced. For speech synthesis, keep stage directions separate from quoted words to be spoken, request a plausible emotional progression and physically coherent delivery, then test candidate audio for intelligibility, instruction fit, authenticity, and event placement. Best-of-N selection is a possible application, not a validated property of these text challenges. If a system needs two scenes with the same voice, the source AC/ACCC specification describes a CUT TO: contrast while holding vocal identity stable.

Relationship to the Voice Acting Arena

No Arena challenge instructions, scene contexts, spoken scripts, audio, votes, or private evaluation split are shipped here. The Arena tasks are held back to reduce direct benchmark gaming; this public dataset has a related but broader construction space and is meant for training, inspiration, and internal evaluation. It is not guaranteed semantically disjoint from every Arena idea: an audit of the 374-task internal Arena snapshot found 11 exact generic-title overlaps with this already-public challenge pool, but zero exact matches between an Arena scene context, performance direction, or spoken script and a published instruction. The private evaluation set and this public pool should still be treated as potentially related for contamination-sensitive comparisons. We disclose aggregate overlap only, without publishing the held-out titles.

Rights, attribution, and limits

The source Voice-Acting-Pipeline repository declares Apache-2.0 for its repository material, and this derived packaging retains that declaration for the released challenge text and scripts. This does not license Eric Morris's books or any third-party source material. The Morris-derived tasks are generated inspirations from acting concepts, not a release of the knowledge-base entries or book pages. Users should review source provenance and applicable rights for their own downstream uses, especially for sensitive commercial reuse.

These challenges are largely synthetic and may contain implausible scenes, stereotypes, emotionally intense material, or inaccurate depictions of health and physical states. Scenario taxonomies expand prompt coverage; they do not establish a representative sample of human situations or validated acoustic consequences. The distribution contains derivative families, repeated titles, and concept overlap. No independent human-quality rating or proof of downstream TTS/video benefit accompanies this release.

If you use the dataset, cite the pinned Voice-Acting-Pipeline, acknowledge the acting-method inspiration and upstream taxonomies, and report which configuration and sampling policy you used.