Send one clip your model gets wrong.
Not a dataset spec: the accent, dialect, noise, scene or phrasing your model fails on. We collect, label and license data that fixes it, in any language. Written spec and quote within 24 hours.
What we collect
Long-tail speech and code-switch
Read, prompted and natural conversation audio in the languages big vendors skip: Taiwanese Hokkien, Hakka, Cantonese, Taiwan Mandarin, Southeast Asian, South Asian and African languages, and mixed-language speech.
Request this →Transcription and word alignment
Verbatim transcripts, word timings and labels done twice by native speakers, with a third reviewer on disagreements.
Request this →Egocentric video from Asian homes and workplaces
First-person video of real chores and work (kitchens, shops, workshops, factories) for robot and world-model training, with task labels.
Request this →Multilingual role-play for voice agents (RL / eval)
Trained native speakers play difficult customers on live calls with your voice agent, scored against a written rubric. Built for RL environments and evaluation.
Request this →Private evaluation sets
Held-out test sets in your target languages and conditions that are never published, with a human baseline.
Request this →Licensed archives
News, books, broadcast and photo collections in 40+ languages, licensed from the rights holders through Norynta.
Request this →How it works
- Day 0 Tell us the failure: what input your model gets wrong. One clip or a short description is enough.
- Within 24 hours Written spec and quote: quantity, languages, conditions, annotation, delivery format, licence. Nothing is signed before you see the price.
- Week 1 A sample in your format.
- Week 2 Paid pilot of about 5% of the volume, with a quality report, so problems surface early. If our spec was wrong, we fix it and re-run the pilot at our cost.
- Then weekly Weekly deliveries to your bucket, each with a manifest, consent reference for every file and a quality report (accuracy, WER where relevant, reviewer agreement, rework rate).
Licence and quality
- Default licence: perpetual and non-exclusive, for training, development and evaluation of your models. Exclusive terms are available at a separate price.
- Data commissioned for you is never resold. Stock and licensed sets are marked as such in the quote.
- Every contributor gives informed consent before recording, can withdraw it, and is paid fairly.
- Rejected items are not billed.
- Independent: no AI lab owns any part of Norynta, so your specs and test sets are not shared with your competitors.
Start a request
One form. You will get a written spec and quote by email within 24 hours. No call needed.
Or email data@norynta.com.
FAQ
What do I need to send?
The failure, not a full dataset spec: the accent, dialect, noise, scene or phrasing your model gets wrong, and one example if you have it. We write the spec.
How fast do I get a price?
A written spec and quote within 24 hours of your request, by email. No call needed.
Which languages?
Any language. We are strongest where large vendors are thin: Taiwanese Hokkien, Hakka, Cantonese, Taiwan Mandarin and Southeast Asian languages, plus mixed-language speech.
How is quality checked?
Each item is labelled or reviewed twice by native speakers; a senior reviewer resolves disagreements and every decision is kept with the item. Each delivery ships with a quality report.
Who owns the data?
You get a perpetual licence to use it for your models. Exclusive ownership terms can be quoted.
How do I pay?
The pilot is paid up front; the main volume is invoiced per delivered, accepted unit.