sidclaw

Thoughtful demo. If possible, I'd want to see next is whether the probabilities it returns are calibrated and not just whether the top choice is right. I ran a small experiment on this, compared base and instruct models' calibration on option-letter logits, then fine-tuned with Brier loss vs cross-entropy. Have you looked at a reliability curve for this version?

fair, it isn't fully native. The dictation helper is the only native piece, the UI is a local web page. That was a deliberate tradeoff to get Qwen3-ASR running through MLX on macOS 14+ rather than wait on a Swift port. Better Voice looks good. Worth noting it's closed source and needs macOS 26, where mine is Apache-2.0 and runs on 14. Different bets.

it's just GitHub interpreting the Co-Authored-By lines in the commits. I used Claude Code quite heavily while building this, so it shows up as @claude in the contributor list — it wasn’t an independent contributor that discovered the repo and submitted a PR :)

good point. I haven’t tested MacPorts, but the app currently looks for ffmpeg on PATH, and the Finder launcher doesn’t include MacPorts usual /opt/local/bin. So it may not detect it at the moment. will look into adding proper MacPorts support.

At least English, German, French, Russian, Italian, and Spanish. You can either select them right in the app or let it auto-detect (which it does pretty well). I think more languages are supported, though, just did not try them.

Well, at least one of us respects it :)

Handy is honestly more mature than mine — cross platform, MIT, and you can pick between Whisper and Parakeet. The real difference is just the model family. I run Qwen3-ASR through MLX instead of Whisper/Parakeet. In my own use it holds up way better with background noise and it punctuate noticeably more naturally, Take that as impression and not a measurement. Qwen3-ASR also accept vocabulary hints, which helps a lot with names and jargon. And mine does files and video with SRT export, not only dictation. Downside: Apple Silicon only. Wispr Flow is a different category — it's paid and your audio goes to their cloud. The text formatting is nicer than mine. Mine is local only and Apache-2.0, which for me is the entire point.

way better from my point of view. I frequently have recordings made in noisy environments. Whisper does not do the job well under such conditions. Also, Qwen respects gramma a lot and splits sentences properly.