4TB of voice samples just stolen from 40k AI contractors at Mercor

Re:4TB of voice samples just stolen from 40k AI contractors ... by numpad0 on Wednesday April 29, 2026 @05:55AM (#47946181)

Thanks - that matches my little observations and clears up a few questions I had. I did notice that spectrogram views are almost the same regardless of perceived gender except that the strongest bands and their distributions change, and also that perceived pitch isn't as dynamic as actual frequency shifts of harmonics, but didn't realize that the center frequency moving up had to do with both physical and figurative size. It makes sense.

One thing I've been feeling itchy regarding this domain is that a lot of existing resources are shallow and there aren't many gamified options even though things like rhythm game feels like a perfect fit. I think not a negligible number of people, especially young and inexperienced, are struggling with aggressive or dis-satisfied sounding voices against their intent. Just a laptop app to feedback the error between their intended voice and recognized voices to let them minimize the error feel like a useful thing to me.

[ Parent | on HN ]

Re:4TB of voice samples just stolen from 40k AI contractors ... by nikkaelle on Monday May 04, 2026 @04:29PM (#48014540)

I've seen some people attempt to make a voice gender recognizer through machine learning, but nothing that has actually worked well. I'm not an ML expert, but I expect they're overfitting on just a small subset of accents and specific formants. The one I played with (can't find the link) was just trying to show weight vs resonance, and it got very confused by my voice. I am never misgendered from my voice but I (intentionally) have a slightly deeper femme voice than what it had been trained on.

IMO, the best way to learn to control your voice is to learn to hear variations in size, weight, pitch, open quotient, prosody, etc. From there you can become your own coach so you're not focused on an app while having a real world conversation with someone. Would a gamified version help? Possibly.

(sorry for the delay, I have no idea if you'll see this. I wrote it last week but apparently never pressed the submit button??)

[ Parent | on HN ]