# The Agentic Data Company

> Conversational data at frontier scale for the labs training the next era of speech models.

## What we do

- We record real people having real conversations and license the resulting corpora to the AI labs and research teams training speech models.
- Each speaker's microphone is captured locally as uncompressed 16-bit PCM and delivered as its own track, so both sides arrive as independent signals on one timeline — overlap, interruptions and turn-taking intact.
- Every hour is consented, licensed and paid for at the source. Speaker identifiers are pseudonymous; names, contact details and account identifiers are never delivered.
- Transcripts are word-level ASR (Deepgram Nova-3) and are not human-verified. Demographics are self-reported at registration and are never inferred from audio.

## Open Yap 1K — published, free

> Channel-separated English natural two-speaker conversations

- 1,000.245 hours across 1,602 conversations and 239 speakers, averaging 37.5 minutes each.
- 48 kHz, 16-bit PCM, one FLAC or WAV file per speaker, with word-level transcripts and per-conversation metrics.
- Free for commercial and research use under the Open Yap 1K Data Use Agreement. Redistribution, re-identification and voice cloning are prohibited.
- Version 1.0.
- [Dataset page](https://theagenticdatacompany.com/open-yap-1k): the full documented release, with measured audio and conversation metrics.
- [Request access](https://theagenticdatacompany.com/open-yap-1k/request): the full corpus, delivered after accepting the agreement.
- [Data Use Agreement](https://theagenticdatacompany.com/open-yap-1k/license): the agreement in full.
- [Sample on Hugging Face](https://huggingface.co/datasets/TheAgenticDataCompany/open-yap-1k): a hand-picked sample under CC BY 4.0 — NOT the full corpus.
- [Docs, schema and tooling](https://github.com/The-Agentic-Data-Co/open-yap-1k): the public repository. No audio.

## Licensed datasets

Not published. Available to buyers under a commercial licence, by email.

- [Task-oriented calls](https://theagenticdatacompany.com/datasets/tool-calling): Scripted voice-agent calls, with mock tools fired at realistic latencies.
- [Accented English](https://theagenticdatacompany.com/datasets/accented-english): Two-speaker English conversations across many accents and first languages.
- [Multilingual conversations](https://theagenticdatacompany.com/datasets/multilingual-conversations): Unscripted two-speaker conversations in the pair’s own shared language.
- [Hinglish, Human Transcripts](https://theagenticdatacompany.com/datasets/human-transcripts-hinglish): Dual-channel Hinglish natural two-speaker conversations with human transcriptions.
- [Multi-party conversations](https://theagenticdatacompany.com/datasets/multi-party-conversations): Unscripted conversations between three to five people, channel-separated.
- [Turn-taking evals](https://theagenticdatacompany.com/datasets/turn-taking-evals): Hand-labelled turn events on unscripted conversations, held out for evaluation.

## Research

Every post is listed at https://theagenticdatacompany.com/research.

- [OpenYap Turn](https://theagenticdatacompany.com/research/turn-taking): a turn-taking model trained only on the corpus's conversations between friends and family; no. 1 on TurnBench end of turn and no. 2 on interruption.

## Legal

- [Terms of use](https://theagenticdatacompany.com/terms): the rules for using this site and the sample audio published on it.
- [Privacy notice](https://theagenticdatacompany.com/privacy): what this site collects from visitors, who processes it, and how to have it changed or deleted.
- [Cookie policy](https://theagenticdatacompany.com/cookies): every cookie this site sets, what it is for and how long it lasts.

## Contact

- Email: matias@theagenticdatacompany.com
- Discord: https://discord.gg/PBvuZQfgw9
- GitHub: https://github.com/The-Agentic-Data-Co
- Hugging Face: https://huggingface.co/TheAgenticDataCompany
- LinkedIn: https://www.linkedin.com/company/theagenticdataco/