Matias Voldby Drejer

@agenticdataco

[Matias Voldby Drejer @matiasdrejer]

@CVestergaard_ Woof

[Matias Voldby Drejer @matiasdrejer]

@nityanandmathur See you there!

[Matias Voldby Drejer @matiasdrejer]

Hello Sydney & Interspeech ✌️

[Midcentury @MidcenturyAI]

Introducing Midcentury.

We’re building the data and simulation infra for physical AI.

Today, we’re coming out of stealth with a $15M Series Seed to scale robotics beyond polished demos.

We’re already supporting frontier labs with:

→ The world’s largest egocentric dataset: 2M+ hours, 50+ environments, 20,000+ tasks

→ Matrix: a frontier simulation platform scaled with our real-world data to evaluate and post-train policies at scale

[Matias Voldby Drejer @matiasdrejer]

@DavidStalmarck @Gustav_Hugod @Atech_dev i swear to god @Gustav_Hugod is @michaelreeves's Swedish embodiment

[Matias Voldby Drejer @matiasdrejer]

@gustavsogard Gopally 😎

[Matias Voldby Drejer @matiasdrejer]

for all my Attio folks out there - created this to sync LinkedIn convos to Attio

https://t.co/IMyCZk1fuj

[Matias Voldby Drejer @matiasdrejer]

@MathiasHeide 🫡 wagmi

[David Stålmarck @DavidStalmarck]

You can now build hardware, you don't even need to wire anything. https://t.co/hLr7BnsKtO

[Gustav Hugod @Gustav_Hugod]

Open source atech duck coming to a repo near you soon👀👀 @Atech_dev

[David Stålmarck @DavidStalmarck]

You can now build hardware, wiring is no longer needed.

And yes, the modules are actually real. ;) https://t.co/hLr7BnsKtO

[Matias Voldby Drejer @matiasdrejer]

@abidlabs @matias Thanks @abidlabs! We looked into gated datasets - the main blockers were per-recipient canaries and wanted a bit more control over the card design. Might revisit and publish the full dataset on HF!

[Matias Voldby Drejer @matiasdrejer]

1,000 hours of natural conversations. Free for commercial use.

Download the full dataset here: https://t.co/ON2PNIF3wd

[Matias Voldby Drejer @matiasdrejer]

@mayanayyar1 @ycombinator @frontrunnerai 🗣️🎤

[Matias Voldby Drejer @matiasdrejer]

we're back in SF. this time as part of @ycombinator F26.

https://t.co/0k1hjeDR5J

[Matias Voldby Drejer @matiasdrejer]

thanks for sharing guys. check out the full dataset here: https://t.co/ON2PNIFBlL

[ModelScope @ModelScope2022]

Open Yap 1K captures 1,000 hours of natural two-speaker English speech.
🤖 https://t.co/XuGT4bNBDh

🎙️ Friends, families, partners, and colleagues talk freely. Interruptions, overlaps, backchannels, pauses, and laughter stay intact.
🔊 Each speaker has a separate 48 kHz track on the same timeline, with transcripts and word-level timing.
🌍 The full corpus contains 1,602 conversations from 239 speakers across 34 countries. Built for speech-to-speech, expressive TTS, ASR, and audio understanding.
📜 Sample: CC BY 4.0. Full corpus: free for approved commercial and research use under the Open Yap 1K Data Use Agreement.

[Matias Voldby Drejer @matiasdrejer]

@DavidStalmarck it would indeed, something like this?

[backchannel @heybackchannel]

To teach AI full-duplex conversation, let it hear the cues.

Open Yap 1K brings 1,000 hours of English two-person conversation for training and evaluating full-duplex voice models.

Speakers talk freely with people they already know, without assigned topics—with interruptions, overlapping laughter, and the little “mhm” that lets a story continue.

Recording — 239 speakers, recorded at 48 kHz, with each speaker captured on an independent track sharing the same timeline.
Overlap — the team reports median overlap of 8.3% of voiced time, reaching 20.9% at the 95th percentile.
Transcripts — Deepgram Nova-3, with word-level timestamps; not human-verified.

Full-duplex AI needs to learn when to join in, when a brief acknowledgment is enough, and when to keep listening.

Access: An 8.9-hour sample is directly downloadable under CC BY 4.0. The full 1,000 hours are free for commercial and research use, subject to application, approval, and a Data Use Agreement.

Credit to @CVestergaard_ and @matiasdrejer at @agenticdataco for the release. Also, love the new company logo—an approving “mhm” from us.

Listen and explore the sample:
https://t.co/dDpDYfUUpS

Project and full-dataset access:
https://t.co/cY4gmyaQeP

#BackchannelSignals

[Matias Voldby Drejer @matiasdrejer]

Open Yap 1K paper now on Hugging Face

https://t.co/JI1SsB1fMw

[Matias Voldby Drejer @matiasdrejer]

This is a serious alternative to speech corpora such as Fisher, Switchboard, and CANDOR. Not only on objective quality measures (48kHz, dual-channel etc), and on the human elements of speech. The interruptions, overlaps, laughter, backchannels and variance in prosody.

[Matias Voldby Drejer @matiasdrejer]

We designed it specifically for full-duplex and interaction models in collab with researchers to capture how people speak together naturally in real environments. Not strangers performing prompted conversations.