How to Transcribe Audio Files on Mac Offline for Free (and When to Upgrade)

The honest answer, up front

If you want to transcribe audio to text on a Mac offline for free in 2026, your realistic options are three: install OpenAI's open-source Whisper on the command line, use a free GUI wrapper like MacWhisper's free tier or Buzz, or write a tiny Swift script that calls Apple's built-in Speech framework. All three keep audio on your machine. All three cost nothing except disk space for the model and electricity to run inference. None of them, as shipped, give you the full package most professionals actually want — live meeting capture, speaker labels, searchable archives, and tidy summaries. That last gap is where you decide whether to stay on free tooling or move to something purpose-built.

This guide walks through the free stack first, benchmarks what each option is good at, then draws the line at which it starts making more sense to upgrade. Every recommendation here keeps the audio on your device — the whole point of transcribing locally is that no vendor ends up with a copy.

Why free, offline Mac transcription is suddenly realistic

Three things changed between 2022 and today. First, OpenAI released Whisper as open source in September 2022, and the Whisper GitHub repo confirms that both the code and the model weights are published under the permissive MIT License. That means you can download the weights, run them on your own Mac, and ship transcripts without ever touching an OpenAI API key. The project's MIT license file explicitly allows use, copy, modify, merge, and distribution — including commercially.

Second, Apple Silicon arrived. Whisper and its derivatives now run well on an M-series Mac without a dedicated GPU. The OpenAI Whisper discussions on GitHub make clear that the small and medium models run locally with no API key required, and community implementations like whisper.cpp add Core ML + Metal acceleration to make inference several times faster on the same silicon.

Third, Apple itself leaned into on-device speech. iOS 26 shipped SpeechAnalyzer, a new modular framework whose Apple Developer documentation describes as fully on-device and tuned for long-form audio like lectures and meetings — and benchmarks suggest it is roughly twice as fast as Whisper Large V3 Turbo on equivalent tasks. The upshot: the gap between "free + local" and "commercial cloud transcription" is now narrow enough that most professionals should seriously default to local.

What macOS can and cannot do out of the box

A common misconception is that macOS has a built-in "transcribe this file" feature. It doesn't. The free-ways guide published by Resonant is blunt about it: macOS Dictation handles live speech into a text field, but there is no built-in way to feed Dictation an audio file and get a transcript back. For file transcription you need a third-party tool.

If you are a developer, you can get further with Apple's Speech framework. According to Picovoice's 2026 iOS speech recognition guide, SFSpeechRecognizer does support file input via SFSpeechURLRecognitionRequest, and setting requiresOnDeviceRecognition = true forces local inference. The catch: SFSpeechRecognizer imposes a hard one-minute limit per recognition session and a rate limit of 1,000 requests per device per hour — which Picovoice notes "rules it out for typical voice memos and meeting recordings." iOS 26's new SpeechAnalyzer removes the cap, but that is iOS 26+ only.

That is why the realistic Mac stack is Whisper-based.

The three free routes, compared

Every tool below runs 100% on your Mac after a one-time model download. Pick based on how much terminal time you want to spend.

Route 1: OpenAI Whisper on the command line

This is the canonical "it just works" route if you are comfortable in a terminal. The side-by-side setup from Resonant's 2026 Mac guide is three Homebrew / pip commands:

1. Install Python 3.8+ via Homebrew: brew install python
2. Install Whisper: pip install openai-whisper
3. Install ffmpeg: brew install ffmpeg
4. Transcribe a file: whisper audio.mp3 --model medium

Keep the network on for the first run so Whisper can download the model, then disconnect it to verify the "offline" claim. A step-by-step verification method documented by Voicescriber's Whisper offline guide is simple: install a local Whisper engine, download the base model while connected, transcribe one test file, then disconnect the network and process a second file. If the second transcript completes, your core transcription path is genuinely offline.

Whisper comes in six model sizes. Larger models are more accurate; smaller models are faster. For long recordings on Apple Silicon, the faster-whisper setup guide recommends large-v3-turbo as the sweet spot — "faster-whisper with large-v3-turbo is the setup for most."

Route 2: whisper.cpp (best for Apple Silicon and CPU-only Macs)

Whisper's Python implementation is excellent, but it brings a PyTorch dependency chain that older Macs don't love. whisper.cpp solves that by rewriting the inference loop in C++ with no heavy dependencies — the project installs via Homebrew on Mac and is specifically called out as the go-to "when you don't have a capable GPU." An open-source wrapper around whisper.cpp documented by Transcribe_Mac on GitHub bundles Core ML + Metal acceleration into a self-contained .app, downloads the Whisper model automatically on first launch (around 650MB, cached), and supports m4a, mp3, wav, aiff, caf, flac, mp4, mov, m4v, aac, 3gp, amr, au, ogg, oga, opus, and webm — essentially anything you'll throw at it.

Verifying it's really offline

Download the model once, then enable Airplane Mode and process a second file. Multiple local-transcription guides including Whisper Notes' 2026 transcription guide describe exactly this check: "Skeptical about 'offline'? Turn on airplane mode first."

Route 3: Free GUI apps — MacWhisper, Buzz, Aiko

Not everyone wants to live in a terminal. The 2026 roundup from Sotto highlights that both Aiko (free) and MacWhisper's free tier transcribe files on-device with a proper interface — drag in a file, pick a model, get a transcript. The 2026 best-offline-apps guide from Voicescriber adds that MacWhisper is "a native macOS app that can run OpenAI's Whisper models and Nvidia Parakeet directly on your machine, so audio is not uploaded when you choose a local model," with a permanent free tier for local transcription.

Buzz: the cross-platform free pick

If you share workflows with Windows or Linux colleagues, Voicescriber calls Buzz "the best free/open-source desktop option" — it transcribes and translates audio offline on Mac, Windows, and Linux using Whisper-based local backends, with file import, microphone transcription, TXT/SRT/VTT export, speaker identification, and speech separation.

How accurate are the free models in practice?

"Modern local models are excellent," Sotto's offline guide concludes. "Whisper Large V3 Turbo and NVIDIA Parakeet v2/v3 rival or beat cloud services for clear recordings, and you can always re-run a file through a bigger model if the first pass isn't good enough." For most meeting audio captured with a decent mic, that is good enough to publish.

Route 4 (advanced): Apple's own Speech framework

If you are building your own tool, Apple's Speech framework is right there. Apple's WWDC23 "Customize on-device speech recognition" session introduced language-model customization for SFSpeechRecognizer beginning in iOS 17 — you can tailor the recognizer to domain vocabulary and improve accuracy without ever contacting Apple servers. The key flag, documented in multiple deep dives including Kamil Tustanowski's Speech framework tutorial, is recognitionRequest.requiresOnDeviceRecognition = true — explicitly set to prevent audio being sent over the network.

The catch for file transcription remains the one-minute cap on SFSpeechRecognizer. If you can target iOS 26 / macOS 26 and up, the new SpeechAnalyzer framework is the right choice. Blake Crosley's comparison of SpeechAnalyzer vs SFSpeechRecognizer describes it as "Apple's modern path: a new on-device model, long-form audio support, automatic language management, low latency for real-time use cases, and a modular architecture" — exactly what long meetings need.

Cloud vs on-device: what you actually give up by going free + local

The main thing you lose by self-hosting Whisper is convenience. The main thing you gain is that audio never leaves your machine, which matters a lot for sensitive material. Here is the honest trade-off:

The Exemplary.ai 2026 Whisper overview is also candid about Whisper's two biggest gaps: "open-source Whisper does not label speakers, and it sometimes writes text that was never spoken." If you need diarization and tight hallucination control, you will want tooling built around Whisper (or around Apple's SpeechAnalyzer), not raw Whisper.

Why free + local matters more in 2026 than it did in 2022

The regulatory and litigation picture around cloud notetakers tightened dramatically this year. On August 13, 2026, Judge Eumi K. Lee of the Northern District of California let the core wiretap, California Invasion of Privacy Act, and Illinois Biometric Information Privacy Act claims against Otter.ai proceed to discovery, holding that plaintiffs plausibly alleged Otter acts as a third-party eavesdropper. The parallel Chamberlain v. Granola case, filed July 30, 2026 in the same district, extends the theory to "invisible" bot-free notetakers that silently capture meetings through a user's device.

None of that touches a workflow where you record a meeting on your Mac and run the audio through local Whisper — because nothing ever leaves your machine, there is no vendor to subpoena and no cloud copy to breach. For a deeper read on the ruling, see our analysis of the Otter third-party eavesdropper ruling and what bot-free meeting recorder actually means.

GDPR adds the regulatory layer. Article 5 of the GDPR requires data minimization — processing only what is necessary. Running Whisper locally is the clearest possible implementation of that principle: the data is never collected by a third party in the first place.

When to stop using free Whisper and upgrade

Free local Whisper is excellent for file transcription. It is not a meeting workflow. Here are the honest signals that it is time to move to a dedicated tool:

• You need to record, not just import. Whisper is a transcription model, not a recorder. If you keep screen-recording your Mac just to capture meeting audio, you've outgrown it.
• You need speaker labels. Open-source Whisper does not diarize. Buzz adds basic diarization, but accuracy varies.
• You need live transcripts during the meeting, not after the fact.
• You want summaries and action items generated the moment the meeting ends.
• You need iPhone + Mac sync so you can start a recording on one and finish on the other.
• You don't want to maintain a Python environment or redownload models when macOS updates.

At that point you're not choosing between "cloud" and "local Whisper" — you're choosing between "wire together five open-source tools" and "install one app that keeps the same on-device guarantee."

How Basil AI solves this

Basil AI is built for exactly the gap between free Whisper and cloud notetakers. It runs on-device on iPhone, iPad, and Mac, with no cloud upload — the same privacy guarantee you get by self-hosting Whisper, but packaged as an app instead of a toolchain.

Under the hood, Basil taps the same Apple on-device speech recognition stack described in Apple's WWDC23 Speech framework session, with the requiresOnDeviceRecognition flag flipped on so audio is never sent to Apple's servers. On iOS 26, the modern SpeechAnalyzer path removes the historical one-minute SFSpeechRecognizer limit so long meetings transcribe end-to-end without stitching. Compared to running Whisper by hand, you get:

• 8-hour continuous on-device recording — no model-size juggling.
• Real-time transcription during the meeting, not after.
• Speaker diarization that Whisper doesn't ship with.
• Action-item extraction and summaries on-device.
• Apple Notes integration via iCloud — see our export workflow guide for Notion, Obsidian, and Apple Notes.
• Zero vendor-held copy to subpoena, breach, or train on.

If you want to keep Whisper for batch file transcription and use Basil for live meetings, both can live on the same Mac. Many power users do exactly that.

Quick decision guide

If you only need to transcribe the occasional interview or voice memo and you're happy in Terminal, pip install Whisper and stop reading. If you want drag-and-drop, download MacWhisper's free tier or Buzz. If you need to actually run meetings — record them, read live transcripts, get summaries, keep everything on-device — a dedicated app like Basil AI is the right upgrade. The common thread across all three paths is the same: your audio stays on your Mac. That is the point.