Trace - No-frills offline meeting transcripts with context
by•
A macOS menu-bar app that turns any conversation into a clean markdown transcript, with a local speech model running entirely on-device.
One global shortcut brings up a small bar at the bottom of your screen. It captures your mic and the system audio as separate tracks, labels who said what, and lets you flag key moments mid-call that sit inline at the right timestamp. No bot joins the call, nothing leaves your Mac, no account, no subscription.


Replies
Trace
Congrats on the launch, this looks super useful (and the fact that files never leave the computer + pricing is a one-time payment, not a subscription, is veryy refreshing). One question on the key moments feature: how easy is it to export or push those specific flagged spots into another note-taking tool like Notion after a call?
Trace
@denitsapenchevavaltchanova super easy! Every recording ends up on disk as a .md and a .json file so you can grep/jq/manually extract it. Key Moments live under their own key in the .json file for exactly this reason.
Here's both files from a transcript that I captured from a YouTube video
Hope that helps!
@alexgodbehere that's super neat!
Really nice idea 👍
How does the local speech model perform in noisy meetings or with multiple speakers overlapping?
Trace
@workout097_collab Good question. The honest answer is I haven't stress-tested either engine in really noisy or heavy-overlap calls. Both Parakeet and Whisper large-v3-turbo are widely used elsewhere so there's a lot of public testing on them, but I can't tell you firsthand how they do in a busy cafe.
A decent mic makes a real difference (and helps with cross-talk from other participants, because your voice is recorded on its own track). The workflow I built Trace around is pasting the transcript into an LLM for a summary afterwards, which is surprisingly good at filling in context even when the transcript is a bit rough in places.
On-device speech is the hard part to get right, and the fact that nothing leaves the Mac is exactly why people end up trusting it. Splitting mic and system audio into separate tracks is a detail most tools skip. I've been through the same on-device vs cloud tradeoff doing local receipt scanning, accuracy is the thing that keeps you up at night.
How's the local model holding up next to the cloud ones?
Trace
@ericlagarda Local vs cloud is definitely a trade-off. In my experience the on-device models aren't anywhere near as good as the frontier cloud ones, but as it stands that's the price we need to pay for privacy. I keep an eye on new releases and hope to pull in better models as they become available, but I also want Trace to stay lightweight and not chew through battery or memory, so it's a balance.
I built this around pasting the transcript into a local LLM afterwards for the summary, and many local models are very good at filling in context and figuring out what was meant, even when the raw transcript is a bit rough (I've found Gemma V4 to be pretty good at this).
On-device transcription with no cloud round-trip is huge for sensitive calls — the use case I see most often is exactly the kind of conversation people don't want flowing through a SaaS pipeline. I host the ModeLoop Podcast on financial modeling and have wrestled with the privacy-vs-quality trade-off for years; the on-device speech models have finally crossed the threshold where local-first is a real choice, not a compromise. Following.