zlacker

The Sky voice was released with the first ChatGPT voices last year in September, so there's no contradiction there unless they asked her on the 1st of September and somehow trained another voice within the few weeks after she said no.

Here's a video that someone posted in October talking to the same Sky voice: https://www.youtube.com/watch?v=SamGnUqaOfU

>>s-lamb+(OP)
> and somehow trained another voice within the few weeks after she said no.

Er, that is totally possible? You act like it's not a machine learning system. You train new stuff in hours or days easily, especially if you have good tooling. Imagine saying this of, say, Stable Diffusion image LoRAs: "this X artist LoRA couldn't be based on X because it was somehow trained within the few weeks after X said no!"

All the timing means is that, in good management & startup fashion, because they needed multiple voices, they had a voice pipeline going, and so could have a bunch of female voices in the pipeline for optionality. And if licensing Johansson didn't work out, you have a plan B (and C, and preferably D). This is big business, you don't do things serially or not have backups: "'hope' is not a plan".

>>gwern+Nn
They could do it but my point is that people are using the September rejection date as evidence for them copying her voice afterwards because it was 7 months before GPT-4o and they aren't aware that the voice has been in the app for 7 months already.