I built a private, on-device meeting scribe for Mac. Optional cloud is opt-in and anonymized first

Thanks :-)

Thanks ! Partially, i used NanoBanana for the first frame, edited it in inkscape to clean it up and convert to vector then made the following frames (4 total)

Thank you :-)

Thanks! TBH, I haven't benchmarked Voice Memos against Thoth on the same recording, so I won't fake a number. I'll do it once I get through a backlog that's gotten a bit long :p and I'll add Apple Speech to my public benchmark while I'm at it. Quick answer tho: one of Thoth's engines is Apple's SpeechAnalyzer, the same one Voice Memos uses, so on your own voice it should be similar. You can also switch to Whisper or Parakeet. Either way, both tools are built for different jobs. Voice Memos records one source, your mic, so on a call it only catches your side. Thoth records your mic and the system audio separately, so it actually captures the other people too, works out who said what, summarizes it, and redacts sensitive stuff before any cloud step.

Thanks :) TBH, i think similar accuracy. MacWhisper & Thoth apps both run the same transcription models, so i don't think there's a meaningful quality gap on transcription itself. I also suppose they do audio processing like me, but i don't know the depth of it. Where Thoth differs isn't accuracy, it's what happens around it: it records your mic and system audio as separate tracks, so the you-versus-them split is deterministic instead of diarized from one mixed stream; and it has an anonymization pass that masks names, IDs, and sensitive terms on-device before any cloud step, with private-enclave (Tinfoil) BYOK if you want a stronger cloud model.

Thank you :-)

Thanks for your comment and for reading my benchmarks :-) On your two questions: 1 Your mic and the system audio are captured as two separate tracks, so the you-versus-them split is deterministic and never confuses. If you need to separate speakers on each stream, Pyannote is what Thoth (and all the industry) uses so i won't expect a massive difference there. Audio capture is what actually matter. For the system stream, not much can be done and Zoom, teams etc... already a good job. On the mic stream, Thoth uses voice isolation and auto gain, to put all chances on user's side for a great audio for transcription and diarization. You can even check the Fourier Transform after the recording to check the quality. Overall, yes, 4-6 overlapping compressed voices on the system stream do make diarization harder, that part is physics, we can only hope Pyannote gets better as time goes by :-) 1-bis Voice profiles for recurring speakers: not yet. What I get now is diarization, anonymous Speaker A/B/C within one recording, that you can rename. Speaker identification, e.g. enrolling a voiceprint once and recognizing that person across future recordings, is on Argmax's roadmap, but it isn't something I ship today so I won't claim it. Definitely adding it once they ship it ! 2 mid-recording copy: yes. The live transcript panel is selectable while recording is running, so you can grab a section and paste it elsewhere without stopping the session. you can even pin the live transcript window so it stays on top of every window. On anonymization, the redaction that runs before a cloud send is four deterministic on-device passes: NER for names, places and orgs (Apple's NaturalLanguage), NSDataDetector for phones, emails, URLs and addresses, your accepted keywords, and renamed speaker labels. On top of that there are two on-device suggestion engines that scan the transcript and propose sensitive terms as chips you accept, one via Apple's Foundation Models framework, one via a verified zero-retention Tinfoil enclave with fail-closed attestation for higher recall. Honest limit for your case: those suggesters find terms, but an accepted term is still a literal match for that value, and there's no way yet to define a standing format rule, so a structured identifier like an MRN or patient ID isn't caught deterministically or carried to the next recording the way your Presidio patterns are. For anything where a single miss isn't acceptable I'd keep Presidio in front of the cloud path. That's the one place a user-definable pattern rule would close the gap, and it's on my list.

Thanks ! :-) Both, user can connect to LM Studio local host and download local models (Phi, Gemma ; Llama ; Qwen)

It uses Whisper, parakeet and Apple Speech, user can choose which ones. They're honestly good, especially if end goal is to extract a summary and action items. I used them at work meetings (dense, technical R&D work, accented english and the results were good, not perfect WER but hnestly good enough for my use case) What really improves is audio capture, eg placing the device where it can capture everyone, or using bluetooth capture tools is what genuinely helps

No difference, all unlocks the same features

it sends the transcript with tokens thet replace sensitive information, like [PERSON_1] instead of Jane Doe, [ORG 1] instead of Microsoft; [PLACE_1] instead of Cupertino... Her's a screenshot as example