Stuff like this is what doctors should use instead of blasting patient data to a AI scribe who knows where. Im pretty sure that is illegal but it still takes place and I am pretty sure one of the reasons for the push to solve FHE...but its obviously way more practical and private to just do the processing locally.
In the US, HIPAA requires companies to sign a BAA to handle any protected information. That's pretty much the only thing stopping data from being shared. That is why, for example, a company website might use Freshpaint or a similar service instead of Google Analytics, as Google will absolutely not sign a BAA for the Analytics platform.
Your data is still collected by a third party analytics company, but they contractually bear all the legal responsibility for maintaining compliance. AI scribe companies are much the same
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.
Question from a naive macports user. To some extent brew is orthogonal to macports. Is there an easy way to link to ffmpeg installed by macports? Thanks.
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.
1. I see it uses python, but I was wondering if it would use mlx-swift instead. Just curious if you considered it.
2. So far, qwen ASR seem usually less well supported than parakeet or whisper so far in most dictation applications (voxtral is similar in that regard). Why do you think that is?
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.
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.
Stuff like this is what doctors should use instead of blasting patient data to a AI scribe who knows where. Im pretty sure that is illegal but it still takes place and I am pretty sure one of the reasons for the push to solve FHE...but its obviously way more practical and private to just do the processing locally.
In the US, HIPAA requires companies to sign a BAA to handle any protected information. That's pretty much the only thing stopping data from being shared. That is why, for example, a company website might use Freshpaint or a similar service instead of Google Analytics, as Google will absolutely not sign a BAA for the Analytics platform.
Your data is still collected by a third party analytics company, but they contractually bear all the legal responsibility for maintaining compliance. AI scribe companies are much the same
How does it compare to models from handy.computer and whisprflow?
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.
That is such a refreshingly fair and honest assessment. Thanks for sharing! To me, that engenders a lot of trust in what you’ve built.
Question from a naive macports user. To some extent brew is orthogonal to macports. Is there an easy way to link to ffmpeg installed by macports? Thanks.
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.
Congrats on the launch!
1. I see it uses python, but I was wondering if it would use mlx-swift instead. Just curious if you considered it.
2. So far, qwen ASR seem usually less well supported than parakeet or whisper so far in most dictation applications (voxtral is similar in that regard). Why do you think that is?
Thanks and good luck with your roadmap!
This is like 2x realtime on M1 vs 60x on fluidaudio with parakeet.
What languages does it 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.
is it better than whisper from oai?
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.
I kinda love that you misspelled 'gramma' when saying this model respects it a lot.
Well, at least one of us respects it :)