The question is not whether it works without WiFi. It's whether anyone but you ever holds the audio.

"Offline" and "local" get used for the same handful of apps, and they are not quite the same claim. Offline is about whether a connection is required at the moment you need the tool — covered in more detail here. Local is a permanent property of how the tool is built: the audio is never sent anywhere, on a good connection or a bad one, this week or in three years. For an interview under embargo, a lecture recorded for a thesis, or a conversation a source asked to keep off any server, that second, stricter promise is the one that actually matters, and it rules out most of the transcription market on architecture rather than on a Wi-Fi icon.

Four tools genuinely hold to it on a Mac today, and they are not interchangeable — one is built around a folder of files you already have, one is built around dictation with meetings added on, one is free and requires you to build the rest yourself, and one is built around a live meeting from the first second of recording. Reading past the shared claim of "local" to what each is actually for saves a wasted afternoon.

MacWhisper — built around files, not a live call

MacWhisper's entire design assumes you already have audio to transcribe: an interview recorded on a separate device, a lecture captured on a phone, a batch of old podcast episodes. Its Pro licence, €64 paid once with lifetime updates included per its own site, adds batch processing across many files at a time, subtitle export in .srt and .vtt, full transcript translation, automatic speaker recognition, and a command-line interface for scripting the whole pipeline. Support runs to over a hundred languages, and export formats extend to Markdown, PDF, HTML and Word. What it does not do is capture a meeting as it happens — there is no live recording mode built around a call in progress, only around audio that already exists or is fed to it after the fact. Full comparison here.

superwhisper — dictation first, meetings folded in

superwhisper's core product is dictation into any Mac app, with meeting recording and transcription added as a feature rather than the whole point — both are included on its free tier, per its own site checked today, alongside support for over a hundred languages and unlimited use of its included models. Its own FAQ is specific that offline models run well on Apple Silicon, with Intel Macs better served by its cloud models instead — worth checking against your own hardware before assuming "local" applies uniformly. It has no meeting summary comparable to a dedicated notetaker's minutes; it hands you a transcript and leaves the write-up to you or to a connected assistant. For someone who dictates constantly and occasionally needs a meeting transcribed, folding both into one app is the actual pitch, separate from how good either half is standalone. Full comparison here.

A self-hosted Whisper script — the only one that costs nothing

Running an open-source Whisper model yourself, either OpenAI's own release or a reimplementation such as whisper.cpp, is free and the transcript quality holds up well against the paid options here, per OpenAI's Whisper repository. It is, by its own documentation, a batch job over a finished audio file — point it at a recording, wait, get text back — with no live capture, no speaker diarization and no summary of any kind built in. Recording both your microphone and a call's system audio in the first place needs a manual audio-routing step, and turning the raw transcript into anything a diarization tool can use is a second manual step layered on top. The honest cost is not the $0 sticker; it is the afternoon spent wiring a Python environment, ffmpeg and an audio router together, plus maintaining that pipeline every time one of the pieces changes. Full comparison here.

Coii AudioNotes — built around the call, not the file afterward

Coii AudioNotes is the one tool here with no batch mode at all, on purpose: it is built around a meeting from the moment it starts rather than a recording handed to it afterward. It captures your microphone and the system audio as two separate tracks, transcribes segment by segment while the call is still happening, tells speakers apart by voice print, and writes the summary — decisions, actions, open questions — with a language model bundled inside the app, so nothing about the write-up depends on reaching a server either. $19, paid once, for three of your own Macs.

That focus is also its limit. It has no batch mode for a folder of old interviews, no subtitle export, and nowhere near MacWhisper's language count for pure transcription work. Full comparison here.

Language support is not the same claim four times over

"Over a hundred languages" appears on more than one of these tools' own sites, and it is worth reading as a starting point rather than a settled score, because coverage of a language and quality within it are different things. MacWhisper and a self-hosted Whisper script both draw on the same underlying open research behind Whisper's language coverage, so their breadth is close to identical even though one is packaged and one is not. superwhisper's language list covers dictation and meetings alike on its free tier. None of the four here publish a per-language accuracy score, so the only reliable check for a language that matters to you specifically is running a real recording through each tool yourself rather than trusting the headline count on its own.

The trade every one of these four makes, and none of the hosted tools do

Local processing moves the computation from somebody else's server onto your own Mac's processor, and that has a real cost the pricing pages for Otter, Notta, Fathom or Granola do not mention because it does not apply to them: on an older Intel Mac, a long recording takes noticeably longer to turn into a transcript, because the CPU that would otherwise sit idle during a call is doing work a server would have done instead. On Apple Silicon that gap mostly closes for a normal-length recording. None of the four tools here can borrow a server's processing power the way Otter, Notta, Fathom or Granola can, because using a server at all is the thing none of them do.

What none of the four try to be

Worth being specific about what stays out of reach across all four, because a category built around "nothing leaves the Mac" gives up whatever a server made possible in the first place. None has a shared team archive another person can search without you sending them a file directly. None has a mobile app that captures audio the way its Mac counterpart does. None pushes a transcript automatically into Slack, Notion or a CRM the way a hosted product with an API key on file can, because maintaining that connection is ongoing work a one-time price or a free script does not fund. Anyone weighing these four against Otter or Notta on those specific rows will find all four lose every one of them, and that is the expected shape of the trade rather than a gap to explain away.

Why local transcription costs more of your Mac's attention, not less

A hosted transcriber sends the audio away and gets text back, so the age of your Mac barely changes the result — a five-year-old laptop and a brand-new one hand the same file to the same server and wait about the same amount of time. All four tools on this page do the opposite: MacWhisper, superwhisper, a self-hosted Whisper script and Coii AudioNotes all run the actual transcription model on the machine you are typing on, which means the machine's own processor is the thing doing the work a server would otherwise be doing. On Apple Silicon that is usually fast enough that a live meeting keeps pace with the conversation and a batch file finishes in a reasonable multiple of its own length. On an older Intel Mac, the same job takes noticeably longer, and running something else demanding — a video export, a large batch job in another app — at the same time will slow the transcription down further, in a way a hosted tool's user would never notice on any machine at all. That trade-off does not show up on any of these four tools' pricing pages, because it is not a price; it is what "nothing leaves the machine" actually costs in return.

Checking the claim yourself, not just trusting the site

Every one of these four tools states plainly, on its own site, that processing happens locally — and that claim is worth spot-checking rather than taking purely on faith, especially for a recording where it actually matters. The simplest check is turning off Wi-Fi entirely and confirming the tool still produces a transcript: MacWhisper and Coii AudioNotes both keep working with no connection at all, per their own documentation, because the model doing the work ships inside the app rather than being downloaded fresh or reached over a connection each time. superwhisper's own FAQ is the most specific about where this holds — Apple Silicon Macs run its offline models well, and Intel Macs are steered toward its cloud models instead, so the same "turn off Wi-Fi" test will behave differently depending on which Mac it runs on. A self-hosted Whisper script needs its model files downloaded once before it can run offline at all, which is worth doing in advance of the day it actually matters rather than discovering the download step while already disconnected.

Matching the tool to the actual audio

The right pick here depends more on what you are pointing it at than on price. A researcher with forty hours of already-recorded field interviews wants MacWhisper's batch processing and export options, not a live meeting app. A writer who dictates all day and occasionally needs a call transcribed is better served by superwhisper folding both into one subscription than by two separate single-purpose tools. A developer comfortable in a terminal, with time to spend rather than money, gets a working transcript from the Whisper script for nothing. Someone whose actual job is showing up to a call, capturing both sides of it, and walking away with finished minutes rather than a raw transcript to write up themselves is the one Coii AudioNotes is built around, and the other three do not attempt that last step at all.

The table

Which to actually try

Point MacWhisper at a folder of recordings you already have before touching anything else on this page — its batch tools go further than any of the other three for that specific job. If dictation is the daily habit and meetings are occasional, superwhisper covers both from one place. If the transcript alone is the entire requirement and you are comfortable maintaining a script, the Whisper route costs nothing but your own time. If the job is a live call, captured as two tracks and turned into minutes the moment it ends, Coii AudioNotes is built around exactly that and nothing else.

Test whichever you choose against the actual audio you have — an old interview, a weekly lecture, a client call — rather than a short sample clip; batch speed, language coverage and speaker separation all behave differently once real length and real crosstalk are involved, and a tool that handled a two-minute test clip cleanly can still stumble on an hour of genuine crosstalk in a way the short test never revealed. Read the offline transcription roundup if connectivity, not just where the audio ends up, is also part of the question, or download Coii AudioNotes directly — thirty days, every feature, no card and no account, with the price only coming up after that.