I've found that for CPU inference the PyTorch-based (non-quantized) version of Pocket TTS actually performs (both speed and quality-wise) better than the ONNX version, even after fiddling with all of the knobs that ONNX provides.
You'd be amazed what exists on the market these days. For example, the pre-terminated InvisiLight fiber cabling is 0.6mm in diameter and has a 2.5mm bend radius. I've personally installed this cabling while making many 90 degree (and sharper in some cases) bends without any issues. That makes it easy to hide and trivial to fit right through doorways and other tight spaces too.
Do you have a link to results confirming this? Kroko does not seem to be on the Open ASR Leaderboard. Parakeet has an average WER of 6.32 across several common datasets.
After having to deal with VM hosts that do GeoIP blocking, which unintentionally blocks Let's Encrypt and others from properly verifying domains via http-01/tls-alpn-01, I settled on a DIY solution that uses CNAME redirects and a custom, minimal DNS server for handling the redirected dns-01 challenges. It's essentially a greatly simplified version of the acme-dns project tailored to my project's needs (and written in node.js instead of Go).
Unfortunately with dns-persist-01 including account information in the DNS record itself, that's a bit of a show stopper for me. If/when account information changes, that means DNS records need changing and getting clients to update their DNS records (for any reason) has long been a pain.
On Linux, this is why I always turn to using abstract sockets when I only need local locking. Only one process can bind and the kernel cleans up automatically on process exit.
You could do the same thing with TCP/UDP, but abstract sockets give you more flexibility in naming with 108 characters vs. being forced to use a 16-bit integer. Also it means you aren't using up a port that could otherwise be used for actual network communication.
Abstract sockets also make for a nice process existence monitoring mechanism since any processes connected to the bound socket are guaranteed to be immediately notified when the process dies.
It's only an alternative if you have a backing swap device. zram does not have this requirement, so (aside from using no compression) it's basically the only solution for some scenarios (e.g. using entire disk(s) for ZFS).