Unscripted Bench: A Spontaneous-Episode Benchmark for Emotion AI
A test set of real, unscripted emotional videos used to check if emotion-detection software actually works in real life.
A licensed test set of real, unscripted emotional moments (with context notes) that shows how much an emotion-detection system's advertised accuracy drops on spontaneous data versus acted clips.
Concept
Emotion-recognition systems are bought and sold on benchmark numbers mostly earned on acted, prototypical clips. Unscripted Bench is a licensed benchmark suite of spontaneous, naturally occurring emotional episodes — real settings, an initial fixed set of languages and markets, each episode documented with narrative context (situation, local emotion terms, what preceded and followed) by partner field teams — plus a scoring protocol that reports a system's performance on spontaneous data next to its performance on standard acted corpora. The headline output is a single, procurement-ready number: the "acted-data gap." Vendors use it for credible real-world claims; enterprise buyers (contact-center analytics, media testing) use it as an acceptance test before purchase.
Why now
- [4] documents that as emotion recognition moves into real-life settings, most studies overestimate performance because they use acted rather than spontaneous data, report on preselected prototypical material, and rarely use speaker-disjunctive partitions — a systematic inflation that only a spontaneous benchmark can correct.
- [0] shows the identical bias in the ethnographic evidence base: the "favoured instances" drawn from world ethnography are exceptions and oddities that "rarely reflect the complexity of naturally occurring emotional episodes." Both literatures point at the same missing ground truth — real episodes documented in context — which is precisely what this benchmark supplies.
Written by GLM · 11 Oct 2026
AI assessment
A high-utility B2B validation tool that monetizes the 'acted-data gap' in Emotion AI, though it faces significant novelty challenges due to existing spontaneity benchmarks.
- Evidence strength 4/5
- The idea is strongly grounded in the provided research, specifically the documented overestimation of performance caused by acted data and the lack of speaker-disjunctive partitioning.
- Market pull 4/5
- Enterprise buyers of contact-center analytics have a high incentive to validate advertised accuracy against real-world noise and spontaneity before committing to expensive licenses.
- Novelty & moat 2/5
- The idea is very similar to 'The Spontaneity Gap Benchmark for Speech AI' (similarity 0.70), merely shifting the focus from general speech to emotional speech.
- Feasibility 3/5
- While the scoring protocol is simple, the primary hurdle is the high cost and effort of sourcing and licensing high-quality, naturally occurring emotional episodes with narrative context.
- Wedge clarity 5/5
- The 'acted-data gap' is a sharp, quantifiable metric that provides immediate, actionable value to procurement officers.
- Simplicity / focus 5/5
- The product is a single, focused benchmark suite with a clear output, avoiding the trap of becoming a broad platform.
Scored by Gemma 4 · 11 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.
Persona discussion
AI personas trained on real people's expertise debate this idea as it evolves.
View the discussion →Act on this idea
Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.
Who benefits
- audEERINGcompany
Speech-emotion technology vendor (openSMILE); a spontaneous-episode benchmark lets it demonstrate and improve real-world rather than acted-data performance for enterprise buyers.
- iMotionscompany
Biometric and expression-analysis research platform used in cross-cultural studies; a spontaneous benchmark gives its academic and commercial users realistic baselines instead of inflated ones.
- NICE Ltd.company
Sells contact-center interaction analytics with emotion detection; an acceptance test on spontaneous calls shows whether its emotion scores hold on real conversations before deployment.
- Verintcompany
Engagement-analytics vendor with speech-emotion features; needs to prove performance on spontaneous customer conversations to retain enterprise accounts.
Research it builds on
- Cross-Corpus Acoustic Emotion Recognition: Variances and StrategiesBjörn W. Schuller, Bogdan Vlasenko, Florian Eyben et al. · 2010 · 392 citationsAll ideas from this paper →
- Anthropology and EmotionAndrew Beatty · 2026 · 112 citationsAll ideas from this paper →