Multi-Speaker Conversation (Audio + Video)
2–6 people talking over each other, every voice on its own track and tied to the face that spoke it
Audio-video generationLip-syncSpeaker attribution
50000+
Hours
3+
Avg. speakers per clip
NEWProduct demos (3rd-person, narrated) · per-speaker dialogue · multi-cam originals
Unaired production footage,
per-speaker tracks
and multi-cam intact
Unaired originals licensed directly from Korean broadcast and live-commerce producers — no crawling, rights clearing, diarization or labeling needed.
Licensed from producers- Non-exclusive · fixed term
- Sample reply in 2 business days
SPK 1 · speakingCAM 1 · 00:12:48Why it’s different
Data you can’t find on the internet
Camera originals,
paired with the broadcast edit
Each camera’s original comes with the broadcast edit
and the matching segments
A mic track per performer
Everyone wore their own mic,
so speaker labels come from the recording, not guesswork
Hand actions with narration
Hosts handle products while explaining them —
action and speech on one timeline
Search the archive in plain language
Indexed by what happens in each scene, not keyword tags. Describe the scene and we’ll find it.
Collections
Samples ready to go
Pick the data you need and request a sample. Open a card to see what’s inside.
Conversational Speech, Speaker-Separated
Natural Korean multi-party dialogue with overlaps, laughter and backchannels, one channel per speaker
Full-duplex voiceDiarizationTTS
50000+
Hours
3+
Overlap ratio
Synchronized Multi-Camera Originals
The same moment from several cameras, time-aligned, paired with the broadcast edit
Multi-view generationEditing modelsWorld models
30000+
Hours
8+
Cameras per scene
Whole-Body Human Motion
Natural motion and interaction from real productions, retargeted to humanoid format
Humanoid controlMotion generation
10000+
Motion hours
Flexible
Robot format
Presenter & Product Demo
Front-facing presenters explaining products, with performer consent for AI training
AvatarProduct video generation
10000+
Hours
2+
Presenters
※ Shown once enough consented footage is secured
Search our archive
We find it in footage we already hold, so samples ship fast
New custom capture
Shot or collected for you, timeline agreed together
Archive
There’s far more beyond these collections
Describe the scene in a sentence and we’ll search our indexed archive and send it as a sample. The five collections above are just a slice of it.
Request sampleLicense terms
License terms
Non-exclusive- Fixed term
- Custom cuts by category, language and hours
About HEIMDEX
About HEIMDEX
HEIMDEX builds on-premises video AI that indexes
broadcast and production archives for Korean media companies.
Backed by AntlerNVIDIA Inception memberClaude Startups program
BACKED BY
Supply model
Clean rights,
a catalog that keeps growing
HEIMDEX indexes archives inside producers’ and broadcasters’ own servers.
New footage keeps joining the catalog without moving the originals, and only licensed clips are delivered.
1
On-premises
STEP 1
Originals stay on the producer’s servers
On-premises2
Indexing · metadata
STEP 2
HEIMDEX curates and labels
Indexing · metadata3
Licensed clips
STEP 3
Only licensed clips are delivered
Licensed clipsAt a glance
How we compare
The four source types buyers most often weigh us against.
| Stock footage licensing | Creator video marketplaces | Crowdsourced custom capture | Public datasets | HEIMDEX | |
|---|---|---|---|---|---|
| Source | Finished edits | YouTube uploads (single cam · compressed) | Staged shoots | Web scrapes | Pre-broadcast production originals |
| Multi-cam · per-speaker tracks | No | No | Separate shoot, extra cost | No | Built in |
| Scene-level metadata | Keyword tags | Barely any | Only what you specify | Varies by dataset | Shot splits · 3-level captions · speaker timeline |
| Finding the scene you need | Keyword search | Channel · category | New shoot (weeks to months) | Filter it yourself | Describe it → we search the archive |
| Rights basis | Platform terms | Individual creator consent | Participant consent | Often unclear | Direct producer contract + per-clip rights record |
| Korean · K-content | Little | Some | Possible but costly | Little | Core focus |
Partner with us
Own a production or broadcast archive?
License your footage as AI training data without moving it off your servers.
How are rights documented?
HEIMDEX licenses footage directly from suppliers who hold the video rights, such as broadcasters and producers. We verify each supplier’s rights and compliance first, then set the scope of use clearly in a legally reviewed license agreement.
Are faces anonymized?
Yes. Faces are blurred or otherwise anonymized by default before delivery. Segments where a specific person’s likeness or IP could be an issue are removed or handled separately, and we agree the scope of processing with you based on your training goals.
What formats do you deliver?
You get video clips cut to just the segments you need, together with labels and metadata. Resolution, clip length, label fields and file formats are set to your spec so they drop straight into your training pipeline.
How fast can I get a sample?
Tell us the data you need, and once we confirm what we hold, we send a sample within 3–5 business days. Check quality and specs on the sample first, then move on to the main contract.
Can you collect to my spec?
Yes, custom builds are available. Tell us the scene type, people and actions, camera perspective (first- or third-person) and environment, and we’ll find it in supplier archives, clean it up and deliver it with the labels you need.
Request a sample
Tell us what you’re missing. We’ll send a cut that fits. 꼭 맞는 샘플을 보내드릴게요.
What happens next
- 1We reply by email within 2 business days — a proposed sample plus 1–2 follow-up questions
- 2Sample delivery: download link + data card + manifest
- 3Within a week of the sample, we confirm format, volume and pricing
sample_hx_talk_000128/
clip_0128_cam1.mp4- clip_0128_cam2.mp4
spk_1.wav
spk_2.wav
spk_3.wav
metadata.json
data_card.pdf
Or tell us more and we’ll pick a closer match