“Offline transcription” can describe the speech-to-text step while the same Mac app still uses the network for setup, updates, or payment. I check those operations separately before deciding whether an app fits a private recording workflow.

I made MinuteFile for Mac, so I can document its boundary precisely. MinuteFile 0.1.8 processes completed recordings and transcripts locally. It still uses the network for optional model downloads, signed update checks and downloads, and checkout.

Check what happens to the recording and transcript

For a completed recording, I start with the two files that contain the sensitive material: the source recording and the transcript made from it.

An offline transcription app should answer these questions plainly:

  • Does speech-to-text run on the Mac or on a remote transcription API?
  • Does the app upload the source recording during import or transcription?
  • Where does the saved transcript and its search index go?
  • Can I review, search, and export the transcript after disconnecting the Mac from the network?

MinuteFile reads recording folders I choose. Once the required model is installed, transcription runs on the Mac. MinuteFile does not upload the recording or transcript to a MotionObj backend or a cloud transcription service.

Saved transcripts, the recording index, People, speaker assignments, bookmarks, and archive search remain on the Mac. Exported files go to a location I choose. The MinuteFile privacy page documents the stored data and network operations.

Separate local work from network work

I do not use “no network” as a blanket description for MinuteFile. The useful boundary is more specific.

Operation Where it happens
Read a completed recording folder On the Mac
Transcribe with an installed model On the Mac
Save and edit the transcript On the Mac
Play the source from a transcript timestamp On the Mac
Search transcript text and filter by an assigned Person On the Mac
Import an existing VTT, SRT, timestamped text, or plain TXT transcript On the Mac
Download an optional transcription or speaker model Uses the model host over the network
Check for and download a signed app update Uses motionobj.com over the network through Sparkle
Buy a licence Opens Stripe Checkout in the browser

This distinction matters when a recording is confidential. A local transcription engine avoids sending the recording to a cloud transcription API, but downloading a model is still network activity. I install any model I need before working somewhere without a connection.

MinuteFile also checks motionobj.com for signed app updates and can download an update through Sparkle. The current MinuteFile appcast identifies version 0.1.8 and includes signed update files. An update request does not turn local transcription into cloud transcription, but it belongs in an honest network inventory.

Checkout is separate again. MinuteFile opens the existing Stripe Payment Link only when I choose to buy a licence. Stripe handles that checkout under Stripe’s privacy policy. The app has no account or subscription; the US$39 perpetual licence is paid once.

The MotionObj website has its own visit and interaction measurement. That website analytics boundary is separate from recording and transcript processing inside MinuteFile. I would not turn “content stays local” into an unsupported claim that every related website and service makes no network requests or performs no measurement.

Check whether the software matches the recording job

Many current Mac transcription results are built around microphone dictation, live meeting capture, bots, or generated summaries. Those may be useful, but they answer a different question.

MinuteFile starts with a completed recording folder or an existing transcript. It does not capture a live meeting, join as a bot, or generate a cloud summary. If I already have an audio file and want the import steps, I use the separate local audio-file transcription workflow. For a finished Zoom recording, I use the Zoom recording workflow.

MinuteFile can also bring an existing VTT, SRT, timestamped text, or plain TXT file into the archive without transcribing the audio again. Recognised timestamps stay connected to local playback. Plain TXT remains editable and searchable, but it stays untimed because the source contains no timing information.

The archive matters as much as the initial conversion. A transcription tool can run a local speech model and still make the saved text hard to inspect, correct, or move. I check whether I can reopen an old recording, play the source beside the transcript, find an exact phrase, and export a normal file without depending on an online account.

MinuteFile keeps the recording reference and editable transcript together in its local archive. It can export TXT, Markdown, or HTML copies. I still keep the original recording and exported copies in my usual backup system; the app archive is not the only copy of material I need to retain.

Review speaker handling instead of trusting the label

Local processing says where computation happens. It does not promise perfect transcript accuracy or automatic knowledge of each speaker.

MinuteFile does not identify voices or infer names. I can keep a meeting-only speaker label, or create a reusable Person and assign that Person manually. Archive search is literal, and I can filter matching text by a Person I assigned.

When one transcript turn contains two speakers, Split Turn keeps the original known timestamp on the first block. The new derived block is marked untimed. MinuteFile does not invent a timestamp or guess who spoke.

I still listen to important passages and correct names, technical terms, punctuation, and speaker labels. Local transcription changes the data path, not the need to review text before publishing or treating it as an official record.

Test offline transcription software on the Mac

I use a short, repeatable evaluation before trusting any transcription app with a private archive:

  1. Install the transcription model while the Mac is online, if the app does not bundle it.
  2. Disconnect the Mac from the network.
  3. Import a completed test recording and run transcription.
  4. Quit and reopen the app, then review and edit the saved transcript.
  5. Search for a literal phrase, apply any local person filter, and export a copy.
  6. Reconnect only when testing model downloads, update checks, or checkout.

That test separates the recording path from supporting network operations. It also exposes whether “offline” applies only to speech-to-text or to the archive workflow I plan to use.

Inspect the local archive after restarting

I do not stop the test when the first transcript appears. I quit the app, reopen it without a connection, and inspect the same recording again. A usable local archive should retain the saved transcript, speaker corrections, and bookmarks. I rerun the same literal search and confirm that the matching archive results still appear.

I also move an exported copy to a normal folder and open it in another app. That confirms I have a file I can back up or use outside the transcription software. It does not prove every internal implementation detail, but it verifies the workflow I depend on.

Record every operation that needs the network

I write down the network exceptions instead of hiding them under one “offline” badge. For MinuteFile, the list is optional model downloads, signed update checks and downloads, and Stripe Checkout. Recording import, transcription with an installed model, transcript editing, literal search, person filtering, and export stay local.

I evaluate accuracy with my own recordings and a model offered by the app. I do not infer perfect accuracy, automatic speaker identity, or broad language support from the word “offline.” Those are separate product questions and need separate evidence.

MinuteFile includes three full trial transcriptions, so I can run the recording, editing, and archive checks before buying. I keep the original recording and exported transcript in my normal backup system, and I make sure I have permission to keep and process the recording.