Regarding the TTS specifically, I remember looking into TorToiSeTTS back when this stuff was first coming out. You can generate ElevenLabs quality audio with it, but it's insanely slow. In fact, when I was looking into it, it seemed like ElevenLabs may have been using a (much faster at the time) version of TorToiSe TTS, given the output is so similar.
According to the linked Github page, they seem to have solved the speed issues now, so it might be worth looking into. Of course, the other commenters have provided solutions that are pre-integrated into the LLM, but if you're just looking for TTS this could be worth checking out. Also worth noting that this requires an NVIDIA GPU.
I always hear about people saying "my subscriptions aren't showing up" and I don't get it. They're right there, sorted in reverse chronological order. I've literally never experienced this problem.
Was going to say this same thing to back you up. We can't both defend this "obvious" study by saying "it's good to have data to back it up!" And then simultaneously argue against having data because it's "too obvious".
I completely agree, a study like this is as good as worthless without disclosing the list of words, or the methodology used for testing words (if they are stored in a latent space rather than a list, for example).
Honestly, before now, this is what I had implicitly assumed was going on. I'm not American, and I had blocked most of this trash when it was happening, but my impression was that these ignorant people were just "coming out of the woodwork" so to speak.
Of course, fear and hate are learned behaviors, but it's interesting to me to see that these people were actually being quantifiably radicalized by that website in particular, and they weren't simply bringing views to the platform that were being established elsewhere.
Of course, not to say the data isn't also important though. It's very possible that we're missing something crucial regarding how the brain functions, despite everything we know so far. The more data we have, the better we can build/test these more streamlined models.
[Edit- I'm blind, the definition I give below does include injury. However, I stand by the fact the word has changed over time, and there is at least some value in following the "old" definition.]
Per Merriam-Webster:
1: to kill or severely injure by electric shock
2: to execute (a criminal) by electricity
Now, granted, because the word is used often enough to mean "shocked", there is a "descriptivist" argument to be made that we should accept the new definition (like "literally" meaning "not literally").
While I'm generally in favour of this approach, I think the distinction here being literally life-and-death (especially when used in a workplace context) warrants some push-back against this new definition.
That said, English doesn't have language police, so you're more than free to disagree with my take, haha.
Regarding the TTS specifically, I remember looking into TorToiSeTTS back when this stuff was first coming out. You can generate ElevenLabs quality audio with it, but it's insanely slow. In fact, when I was looking into it, it seemed like ElevenLabs may have been using a (much faster at the time) version of TorToiSe TTS, given the output is so similar.
According to the linked Github page, they seem to have solved the speed issues now, so it might be worth looking into. Of course, the other commenters have provided solutions that are pre-integrated into the LLM, but if you're just looking for TTS this could be worth checking out. Also worth noting that this requires an NVIDIA GPU.