Company profile · 3 min read
Liva AI: consented human voice and video data for model builders
Liva AI collects rights-cleared human voice and video data across languages, accents, emotions and contexts for AI labs and multimodal product teams.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Liva AI supplies human voice and video data for companies building realistic multimodal models. Its buyer is a voice, video or audio lab that needs diverse accents, languages, emotional range and contexts that scraped internet data cannot reliably provide.
What it does
The YC launch says Liva collects real human voice and video in-house through crowdsourcing and partnerships, with consent and rights clearance. Examples include sales calls, multi-channel dialogues, expressive monologues, casual conversations and job interviews. The company says it is already delivering a voice dataset to a lab training expressive foundation models.
The product is data supply and collection operations, not a model API. The YC profile describes Liva as building socially intelligent AI, starting with voice. The public homepage was not readable in the checked packet, so current catalog, pricing, delivery formats and customer references are not established beyond the launch material.
Why I’d look closer
The advantage is authentic context. Models that need human-like interaction require more than clean sentences: they need turn-taking, emotion, accents, interruptions and social situations. Ashley Mo’s public background includes Caltech CS, MIT health-audio data collection and biosensor work; Aoi Otani is described as a Harvard bio/CS researcher with representation-learning and diffusion-model publications.
The tradeoff is consent and representativeness. “Rights-cleared” must cover the recording, intended model use, derivative data, voice likeness, retention and deletion. A large dataset can still overrepresent certain accents, cultures or scripted performances, and sensitive conversations need a higher bar than ordinary speech.
What I’d ask
How are participants recruited, paid, consented and allowed to withdraw? What rights does a customer receive, and can the data be used for voice cloning or only evaluation? How are accents, identities, health references and minors handled? Can the buyer audit provenance, labels, exclusions and deletion across vendors and storage?
My editorial take
Shortlist Liva if your model’s bottleneck is socially rich, rights-cleared audio or video rather than raw volume. Start with a narrowly defined collection brief and require participant-level consent and provenance evidence before training. The differentiator is not “real data” as a slogan; it is whether the data remains usable, lawful and representative after deployment.
Quick facts
| Field | Sourced detail |
|---|---|
| Product | Consented human voice/video data collection and delivery |
| Buyer | Voice, video, audio and multimodal AI labs and product teams |
| Contexts named | Accents, emotions, sales calls, dialogues, monologues, interviews |
| Pricing | Not published in the checked pages |
| Main question | Can the team prove consent, provenance and diversity for every recording? |
Sources checked
| Source | Checked |
|---|---|
| YC company profile | 2026-09-19 |
| Liva AI YC launch | 2026-09-19 |
| Liva AI website | 2026-09-19 |
Cohort context
Liva AI is listed in Summer 2025. In our 2026-09-18 directory snapshot, 112 of 166 listed companies in that cohort have YC’s primary industry label B2B (67.5%). This is a current-directory comparison, not an original intake count or a performance ranking. Nine-cohort dataset.
Public website snapshot
Observed 2026-09-19T16:18:09.383Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.
| Signal | Homepage observation |
|---|---|
| Product description metadata | Not observed in this response |
| Canonical link | Not observed in this response |
| H1 or H2 heading | Not observed in this response |
| Typed structured data | Not observed in this response |
| Docs/developer link | Not observed in this response |
| Pricing link | Not observed in this response |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |
Public observations · Collection method. Missing links here do not establish that a capability or file is absent elsewhere.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗