Liva AI

What do they actually do

Liva AI builds and sells curated, rights‑cleared voice and video datasets recorded from real people under explicit consent. They design and run the capture (accents, emotions, contexts) and deliver audio/video files with metadata and provenance, emphasizing original recording rather than synthetic or scraped sources (YC company page · theliva.ai).

Customers engage via a contact form or direct intro, scope needs (demographics, scenarios, recording specs), and Liva runs or supervises production, clears rights, and hands off training‑ready data plus consent documentation. Public materials indicate bespoke dataset access and custom projects rather than a self‑serve catalog today, and they note at least one delivered dataset to a lab training expressive voice models (YC company page · theliva.ai).

Who are their target customer(s)

• AI research and model-building teams at labs training voice or multimodal foundation models: They need large volumes of realistic, consented human recordings across accents, emotions, and interaction types; public or scraped data lacks coverage and clear rights, which degrades model quality and adds legal risk (YC launch post · theliva.ai).
• Product teams at startups building voice/video features (avatars, assistants, dubbing): They require domain- and emotion-specific samples to fine‑tune models but lack the production workflows and legal clearance processes to source this data themselves (theliva.ai).
• Enterprise ML teams in regulated domains (contact centers, healthcare, education): They need scenario‑specific, consented recordings with audit‑ready documentation; sourcing compliant, high‑quality data at scale is slow and risky without a specialized vendor (theliva.ai).
• Data procurement and legal managers at AI companies and labs: They must avoid IP/privacy exposure and require explicit consent records, usage rights, and provenance trails from vendors—artifacts many dataset sellers don’t provide (YC launch post · theliva.ai).
• Academic researchers and small labs studying human expression or dialogue: They want richly labeled, context‑specific recordings, but public datasets are often low‑quality or missing needed interaction types, and running compliant collections is expensive and time‑consuming (theliva.ai).

How would they acquire their first 10, 50, and 100 customers

• First 10: Leverage YC and founder networks for warm intros to AI labs and model teams; run fast, tightly scoped pilot datasets that include full consent/provenance packs to convert pilots into references (YC launch post · theliva.ai).
• First 50: Package early work into 2–3 standardized dataset “packs” (accents, multi‑party calls, emotional monologues) and onboard production partners to increase throughput; target startups building voice/video features and demo at conferences to drive inbound (theliva.ai · YC launch post).
• First 100: Stand up procurement‑ready offerings (standard contracts, audit trails, provenance dashboards) and add channel partners (AI vendor marketplaces, procurement platforms, production houses); hire a small BD/sales team and launch a basic self‑serve catalog or API for smaller buyers.

What is the rough total addressable market

Top-down context:

Industry reports estimate the AI training‑dataset market at about $2.9–3.2B in 2024, with projections into the mid‑teens of billions by the early 2030s (Fortune Business Insights · ResearchAndMarkets). Image/video is cited as a large share (≈40%), and separate estimates put audio datasets around ~$1.4B in 2024 (Grand View Research · Market Intelo).

Bottom-up calculation:

Applying a ~40% share to the ~$2.9–3.2B AI training‑dataset market implies ~$1.1–1.3B for image/video, and adding audio dataset estimates (~$1.4B) suggests a current audio+video TAM around $1–2B for paid, rights‑cleared training data, growing with the overall market (Fortune Business Insights · Grand View Research · Market Intelo).

Assumptions:

• Only includes paid, third‑party datasets with consent/provenance suitable for training; excludes internal data and open web‑scraped corpora.
• Shares for audio vs. image/video are approximate and vary by source, but are used to bound a reasonable range.
• Demand for rights‑cleared human expression data scales roughly with broader model training activity and dataset spend.

Who are some of their notable competitors

• Defined.ai: Provider of speech/audio datasets and custom data collection with consent and QA; competes directly on curated speech corpora and provenance controls.
• Appen: Large data vendor offering audio/video data collection and annotation; strong enterprise relationships and global crowd operations.
• Scale AI: Broad AI data platform with collection/annotation services for audio/video; notable for serving top labs and offering enterprise workflows.
• TELUS International AI Data Solutions: Formerly Lionbridge AI; provides global speech/audio and video data collection/labeling with enterprise compliance programs.
• SpeechOcean: Supplier of large speech corpora and custom collection services across many languages; known for catalog depth in speech datasets.