Sep 1, 2026, 5:16 PMMeta continues to be hilarious rolling out Meta Voice Transcribe. Meanwhile, we're all doing transcription inference locally with Whisper.
Christ, I'm happy not to be working there any more.
- Sep 1, 2026, 5:16 PM
Replies
FXBOY4EVAFXBOY4EVA@mastodon.socialSep 1, 2026, 7:09 PM@syn been trying this out with a fairly compact (<1Gb) parakeet model recently (though it can optionally pull and use Whisper. Moonshine, SenseVoice, Paraformer and others).
A little finnicky to set up initially (especially on a desktop system running X11 rather than wayland), but, surprisingly performant and accurate even on relatively low power boxen once it is.
- Sep 1, 2026, 7:15 PM
@FXBOY4EVA I use a different Whisper harness but it's all the same underneath. Don't rate Parakeet. Those other models are new to me though, might give that a pop. Ta!
- FXBOY4EVAFXBOY4EVA@mastodon.socialSep 1, 2026, 7:16 PM
@syn Yeeeerp 👍
- Sep 3, 2026, 2:19 PM
@FXBOY4EVA wow, you weren't kidding about this being annoying to set up.
the homebrew build only has one engine, which is no good when I want to try all the models. not a huge deal, I've built it manually (god bless
cargo).then while the model downloader is turbo fucking annoying and must be interacted with through
readline, i've got the Warp agent interacting with the command and iterating through the available models.so yeah. nothing which is a blocker, just jesus christ more trouble than it needs to be.
- Sep 3, 2026, 2:46 PM
@FXBOY4EVA Fuck me, that was worth it. The Cohere model might be heavy but it's incredibly good, fast, and doesn't even spank the GPU.
- Sep 3, 2026, 2:49 PM
@FXBOY4EVA oh, and another thing. THERE ARE TWO FANS RUNNING ON FULL PELT IN MY ROOM. Cohere doesn't care.
- FXBOY4EVAFXBOY4EVA@mastodon.socialSep 3, 2026, 7:20 PM
@syn Yeah, the input filtering is surprisingly good even still using the Parakeet models, got a whole paragraph cleanly in a very noisy coffee shop on just a built-in laptop mic.