wentw0rth

You are correct with inferring it's a scale, completely.

Politely disagree with "You aren't going to be able to find a set of signatures that keeps out a useful number..."

That's AGEWARDEN's job and the last year of my life has been diving deep into audio data to find those exact signals.

Good call out and thank you for the pushback :)

Love it. We've been playing around with other ideas that are not web related. As the tech improves, less data will be needed to infer age. Strange now, but setting up a microphone at a kiosk or providing a widget-less API for broader use-cases, is not a stretch for piggy-backing. Thank you for the comment.

Good question: regardless of the recording instrument (our lab has many microphones of varying degrees from low > high fidelity), a live person vs. recordings that are played back carry a unique signature.

That would be considered a "replay attack" and recorded audio carries a unique signature: BLOCKED

The AGEWARDEN system is ^^ agnostic and answers a simple question only: Is this human over or under 18?

From my research, each person is incredibly unique :)

> There should exist frequencies impossible to reach for anyone but post-pubescent men.... Of course that would make it teen-age verification at best.

Yes and no. Like geology, the more you dig into audio, the more is uncovered and available.

A child/teen with a deep voice has very unique perturbations that don't exist in other cohorts across feature sets. Adults may share some characteristics, but not at the levels found in specific age groups.

> Your comment brought my mind to voice actors. They easily change their voices. Deep fakes also come to mind.

Voice actors, deep fakes, replay attacks (recordings from a TV/computer/phone) - mitigated. As in my earlier comment where teen voices have a signature, these all have unique signatures as well. The ongoing part is staying up-to-date with the latest synthetic (deep fakes/AI generated) as those are getting better each day.

Post edit/updated, let me track:

> How would we trust your platform to not store voice fingerprints, then?

This is a good question. Other than having it be more expensive to keep this data, and if it were true, destroy the company whose core tenant is literally that... you can't.

If you have suggestions for an independent audit, more than happy to add that to our stack where applicable.

When I say it's hard to not collect data, woof. Any fresh AWS account just LOVES to slurp it up by default.

All good! Posting here to get the opinions and feedback from engineering peers.

> I... told it in a higher then normal voice that I was a kid, and it said I was under 18...

Good. If you're pretending to be a kid, you should be blocked. The launch is for sites that allow adults only (should have made that more clear in the post).

Deep voiced teens are THE HARDEST. What I commented earlier about the amount of data available in a clip of audio, that's real - teens have unique signatures that only show in their cohort.

Thanks for testing!

Lawyers HIGHLY suggested the line; AGEWARDEN is > 95% accurate and on par with others in this space.

If someone told me they had 100% accuracy with inference, I'd call them out.

Thank YOU for calling it out :)

That's the nice thing about audio temporal vs. video, much cheaper.

LOL, it was BEFORE the sketchy times :')

Thanks for the comment/question:

> wouldn't it be far easier to bypass this than even face (live video) verification... or am I missing something?

Missing the experience of learning just how much information is in a clip of audio; replay/synthetic attack mitigation is baked in, and more cost-effective to run vs. video, and less creepy.

The mission: make the process easy (and inexpensive) as possible so that people can be verified at scale, without giving away their personal data and being the product.