# Askable Labs

> Human-data production partner for frontier AI labs. Not a data vendor — there is no shelf dataset. A lab brings a capability goal; Askable co-designs the data spec, recruits the right people to it (representative general population to credential-verified experts, 50+ countries), and produces a consented, validated, schema-conformant dataset: multimodal capture of real people doing real work — spoken reasoning, screen, audio, timestamped transcripts — with identity triple-verified and brief-versioned consent, withdrawable until release. ISO 27001/27701 certified, SOC 2 Type II attested, UK Cyber Essentials certified, GDPR/CCPA compliant. Operating since 2017.

## Core

- [Capabilities](https://labs.askable.com/agents/catalogue.md): the four data types a lab can commission, in the order the site presents them:
  - **Long Horizon Captures**: Agents, SFT, Multi-agent, Tool-use
  - **Conversational Audio**: ASR, STS, TTS, Voice
  - **Evals and Preference**: Reward models, Evals, Preference, Ranking
  - **Physical Skill**: Video models, World models, Embodied

  Each data type is captured to a bespoke brief, with a few example captures listed per type. The full pattern library is internal. Delivery schema, capture moments, and the T01–T06 capture stack: [keys](https://labs.askable.com/agents/catalogue/keys.md). Human page: [/data/](https://labs.askable.com/data/).
- [Panel & expertise](https://labs.askable.com/agents/panel.md): who is on the panel and what they do. ~900,000 participants and experts across 75 countries, 6 continents, and 50+ languages, with a country-by-country coverage table, domain and profession maps, and the expertise gradient. Answers "do they have <this kind of participant>?". Human page: [/experts/](https://labs.askable.com/experts/).
- [Overview](https://labs.askable.com/agents/overview.md): who Askable Labs is, what data it can produce, what makes it different, legitimacy and consent posture, in Q&A format
- [Production method](https://labs.askable.com/agents/method.md): the six-step pipeline (spec co-design → recruit → consent → capture → process → validate), delivery artifacts, audit posture
- [How to engage](https://labs.askable.com/agents/engage.md): spec-led vs outcome-led engagement, capacity and timeline shape, a fill-in data-spec template, contact

## Optional

- [Capabilities catalogue, JSON](https://labs.askable.com/agents/catalogue.json): full-fidelity structured dump of the catalogue for programmatic use
- [Everything in one file](https://labs.askable.com/llms-full.txt): overview, method, full catalogue, and engagement concatenated
- [Experts (human page)](https://labs.askable.com/experts/): the people captured, recruitment, credential floor, and identity verification
- [Research (human page)](https://labs.askable.com/research/): notes and methods on capturing expert work as training material
- [VOICE-H (human page)](https://labs.askable.com/voice-h/): interactive results of VOICE-H, the Labs human evaluation benchmark for TTS models — Elo leaderboard with per-category scores, head-to-head records, price vs performance, and rater demographics
- [Enterprise (human page)](https://labs.askable.com/enterprise/): always-on data, human evals, and fine-tuning data for applied-AI enterprise teams
- [Security detail (human page)](https://labs.askable.com/security/): certifications and compliance posture
- [Contact (human page)](https://labs.askable.com/contact/): partnership inquiries

## Research articles

- [Introducing VOICE-H (updated).](https://labs.askable.com/research/voice-preference-eval/): VOICE-H is a human evaluation benchmark for TTS models: 100 quotes selected from 40,000 real interviews, 17 models against the original human recordings, and the qualitative reasons behind every preference. Updated with 15,217 English comparisons from 901 rater sessions, three new models, and French and Spanish results.
- [VOICE-H expanded.](https://labs.askable.com/research/voice-h-round-two/): Five new TTS models join VOICE-H, taking it to fifteen. Cartesia's Sonic 3.6 takes the top place, Speechify's Simba 3.2 arrives third at a fraction of the price, and the same model can lead in one language and sit mid-table in another.