Guide

Why transcription is slow on your Mac

The same recording can take two minutes or forty on different Macs running the same app. Four things explain almost all of the difference.

30 August 2026

Local transcription speed varies enormously, and not in proportion to how new your Mac is or how much you paid for it. The same one-hour recording might take three minutes on one machine and half an hour on another, with the same app.

Four things account for nearly all of it.

1. Apple Silicon versus Intel #

This is the big one, and it is not a small difference.

Transcription is dominated by matrix arithmetic. Apple Silicon has two things that suit it specifically: a Neural Engine built for exactly that arithmetic, and unified memory, so the model's weights do not have to be copied back and forth between CPU and GPU memory.

An Intel Mac has neither. It has no Neural Engine at all, so the encoder — the expensive half of the model — falls back to the CPU or a discrete GPU across a bus. The result is often several times slower on the same file.

If you are on an Intel Mac, transcription is still perfectly usable; it is just something you start and come back to rather than watch. Nothing in software closes that gap, because the gap is a piece of silicon.

2. Whether the app uses the Neural Engine at all #

Two apps on the same M-series Mac can differ by a large factor, because using the Neural Engine is not automatic.

The model has to be compiled ahead of time into Core ML's format for the encoder to run there. Apps that ship a Core ML encoder alongside the weights get the fast path; apps that do not run everything on GPU or CPU and leave a lot of performance unused.

You cannot see this from outside, but you can feel it: if one local transcription app is dramatically faster than another on your machine, this is usually why.

3. Model size #

Whisper's sizes differ by roughly 3× in parameters at each step, and time scales with them. Going from Base to Large is around twenty times the parameters — on a long file that is the difference between a coffee and an afternoon.

If an app defaults to a large model, or you selected one because bigger sounded better, that alone can explain the wait. The accuracy gain is much smaller than the size difference suggests, and for clean speech it is often imperceptible.

4. How much silence you are paying to transcribe #

A two-hour recording of a ninety-minute meeting contains thirty minutes of nothing. Without voice activity detection, the model processes all of it — decoding silence at the same cost as decoding speech, and producing repeated junk text while it does.

Apps that run VAD first skip those stretches entirely. On recordings with a lot of dead air, this is not a marginal saving.

What actually makes it faster #

In order of effect:

1. Use a smaller model. Base is enough for most clean speech, and it is many times faster than Large.
2. Use an app with a Core ML encoder if you are on Apple Silicon.
3. Trim dead air, or use a tool that detects speech first.
4. Close whatever else is competing. Transcription will use every core it can get. A video export in the background halves your throughput.
5. Plug the laptop in. On battery, macOS throttles sustained work; Low Power Mode does so aggressively.

What does not make it faster #

More RAM, unless you were swapping — the working set is smaller than people assume.

Faster internet. There is no network step in local transcription. If your speed depends on your connection, the work is not happening on your machine.

Restarting it. If a run is slow, restarting begins the same work from zero.

A rough expectation #

On an Apple Silicon Mac with a Base model and a Core ML encoder, transcription runs several times faster than real time — a one-hour recording finishes in a handful of minutes. Our own measurement on a two-hour file was under three minutes end to end.

On an Intel Mac, expect it to take a meaningful fraction of the recording's length. Still worth doing, still free, just not something to watch.

If you are far outside those ranges on Apple Silicon, the likely causes are, in order: a large model, an app not using the Neural Engine, or a lot of other work competing for the machine.

The short version #

Apple Silicon versus Intel is the biggest factor and you cannot fix it in software. After that: use a smaller model, use an app that compiles the encoder for the Neural Engine, skip silence rather than transcribing it, and do not run it alongside a video export on battery power.