Let it run for a couple hours and it got to the F1 to continue or F2 to enter setup, but wouldn’t respond.
O.K. Went to get lunch, about 1.5 hours later, I just walked by and stabbed the F2 key, and whaddya know, it entered setup. But it’s not responding again, or maybe it’s responding v e r y s l o w l y. I have a theory involving a miniature blackhole and time dilation that’s my current leading idea.
It shows the old processor specs, but has the new processors. It shows the old ram specs(1333mhz x8gb) despite having the newer ram(1600mhz x16gb), but has correctly identified which slots are populated. The main header also still shows 192gb in all 24 slots. Probably just an artifact of not getting fully booted and settled???
I’ll add this too, but I think it shows no disks because it hasn’t completely booted up yet.
I’ve had a breakthrough.
Been trying to use the mouse / keyboard on the server itself (USB). I was looking at the virtual console on IDRAC and when I tried the virtual keyboard, it responded. Long story short, clearing out the lifecycle jobs/tasks/whatever seems to have done the trick. It went through, found all the new hardware, set everything back up, but… It just won’t boot now.
Long button mashing story short, apparently getting rough with lifecycle controller can bork Grub. Rather than trying to repair it, I just re-initialized the virtual disk and installed the latest debian (13.3). Anywho, I’m in now, and I’m back to a blank slate.
Funny you should mention that, I’ve just upgraded myself and I’ve got all this RAM just laying around. So if a forum member wants to stuff their system full, lemme know.
192gb ram (24 sticks of 8gb @ 1333mhz) I know it fits / works in a Dell R720, other servers??? There’s 12 sticks of M393B1K70CH0-YHQ5 and 12 sticks of m393b1k70DH0-YH9. The only difference is a slight change in latency, but they’re perfectly compatible.
2x E5-2630 v2 CPUs These are high core count at 2.3ghz, I changed to lower count to gain speed.
2x 750w power supplies that fit the Dell R720. Perfectly good, I just went bigger for GPU support.
First come, first served, but these could be good parts for someone’s homelab (doesn’t have to be A.I.)
I don’t know who’s in / around / near Salisbury, NC, BUT…
There’s an auction for 5 Dell R720 Servers. It’s going down in less than 4 days from right now.
So, A. You could have a literal stack of servers to play with, or B. you could keep one (maybe 2) and ebay the rest. Given that the high bid is currently 75 dollars, it’s something to think about.
Either way, nobody’s spoken up for the extra ram and stuff, so I could hook you up.
Ok this is the number one reason to switch to kokoro vs piper.. piper is just box-o-rocks stupid when it comes to interpreting text sent to it. Kokoro is sooooo much smarter.. it knows how to handle things like “10:27pm”. Just a quick test I did shows there’s still much to be desired, but still just getting the goddamn time correct is fantastic.
its 10:34am and 10-4 good buddy. the date is 2/22/2026 ok?
“ten minus four”.. though I didn’t expect it to understand the nuance of 1970s trucker language. let’s put spaces in.
its 10:34am and 10 - 4 good buddy. the date is 2/22/2026 ok?
Yeah ok. I would still love for it to handle dates correctly but then again “correctly” for who.. for dates like 02/02/2022 you get into that whole month/day order problem, so I get why they might want to steer clear of that one and let the LLM or text author deal with it if they want proper dates read out.
Anyway I’m back on the LLM kick for home assistant now.. just to see if I can make it work better at this point or not.
Just digging in a little more.. the debug for kokoro shows what’s happening to the text.. it’s not “smart” per se, it’s doing some pattern based text transforms before generating speech. I found this in the debug log;
07:14:59 AM | DEBUG | kokoro_v1:261 | Generating audio for text with lang_code ‘a’: ‘its ten thirty-four am and ten - four good buddy. the date is two slash twenty-two slash twenty twen…’
The funny thing about it is that it reads out everything as a statement, even “Are you ok? How are you doing?” is read “are you ok. how are you doing.” .. it’s offputting when an LLM responds with more interesting interactions than the built-in home assistant command processor.
[S1] Hey what time is it?
[S2] it's almost 10:30am
[S1] Wow. Amazing. (laughs)
So.. I guess we go back to Kokoro TTS. Like wtf is that random noise at the end.. youtubers also seem to be experiencing this as well. I guess that’s the difference between a “TTS” and “generative voice” software.. GV is meant to make audio outputs you’ll probably edit, regen until it’s right, then put into media of some kind.. whereas TTS is meant to generate something descent the first and only time, and not have any bullshit, audio bugs, etc.
Let’s also talk about time.. Kokoro TTS is almost (basically) instant.. Chatterbox takes 5-10 seconds on one of my RTX 3060s for even a short sentence.. and Dia takes like 60 seconds on one of my Tesla T4 cards. Out of the stuff I just tried, Kokoro TTS is obviously the most suitable.. even if it completely lacks intonation awareness.
I’ll check out the turbo option but one thing with chatterbox and these other generative voice models that work by conjuring phonics and phomems from nowhere, vs a more traditional TTS system, is that they often generate random noise .. often at the end of generation, so you get the text but then you just get this strange hissing and popping and buzzing noise for like 2 to 15 seconds.. and it’s random. So far I haven’t seen a way to mitigate this.
Probably just an artifact of not getting fully booted and settled???
Dell R720 Servers. It’s going down in less than 4 days from right now.