Parakeet vs Whisper on a Mac: we timed both on the same 15-minute meeting

What we measured

One real 15-minute meeting recording (896 seconds, the call side, mono 16 kHz), on an M2 Max running macOS 14.8.5. Parakeet TDT v3 through FluidAudio 0.15.7, Whisper through WhisperKit 0.9.4, both in release builds with the models already downloaded. Numbers are warm runs; the first run of any model pays a one-time Core ML compile, which for Whisper Turbo was 151 seconds and afterwards 3.5.

The recording is a real meeting, so no transcript text is published here. The only outputs quoted are counts: words, punctuation marks, and how much the engines agreed with each other.

Speed, CPU, and memory

CPU is the share of one core over the transcription, so 114% means Whisper Turbo kept more than a full core busy for its 51 seconds while Parakeet used a twenty-fifth of one. That is the number that decides whether your fans spin up and how much battery a day of meetings costs, and it is where the gap is widest.

Whisper Large v3 is the accuracy flagship and it took six minutes for fifteen minutes of audio. It still beats real time, so it works for recordings, but you would not want to wait for it after every call.

Why the same model can be four times slower

Speed is not a property of the model alone. The same Parakeet build, on the same file and the same Mac, ran roughly four times slower and used several times the memory when it was left to run the way it was set up out of the box. Different chips and different macOS versions land differently, which is why published figures rarely match what you see.

Mac Note Taker tunes this for your Mac by itself on first load, so the table above is what the app actually does rather than a best case. If you are comparing transcription tools, run them on your own machine before believing anyone's number, ours included.

How different are the transcripts?

Without a hand-made reference transcript there is no honest word error rate to quote, so we measured agreement instead: how many words two engines put in the same place. Parakeet against Whisper Turbo agreed on 85.7% of words, against Whisper Large v3 on 87.5%. The two Whisper builds agreed with each other on 92.0%.

That last number is the useful one. Two builds of the same model family disagree on 8% of words on ordinary meeting audio, so a 12-14% spread between Parakeet and Whisper is smaller than it sounds. The disagreements are mostly filler words, false starts, and where one engine dropped a short interjection the other kept.

Parakeet wrote 831 words to Whisper Turbo's 756 on the same audio, with 87 sentence endings and 42 commas against Whisper's 107 and 57. Whisper punctuates a little more densely and drops more crosstalk; Parakeet keeps more of the short overlapping speech, which in a meeting is often someone agreeing or interrupting.

What neither of them does

Neither engine tells you who spoke. Speaker separation is a different model entirely, running beside the transcription, and matching a voice to the same person across meetings is a third step on top of that. Any app that shows you named speakers is doing work these two models do not do.

Which to pick

• Meetings, English or another of Parakeet's 25 European languages: Parakeet. Faster, far cheaper on battery, punctuation good enough to read.
• Audio Parakeet struggles with - heavy accents, poor microphones, a lot of crosstalk: try Whisper Large v3 and accept the wait.
• Languages outside Parakeet's list: Whisper, which covers 99.
• A long backlog of recordings to process: Whisper Turbo or Distil are the sensible Whisper builds; plain Large v3 is 7x slower for a small gain.
• On battery, away from a charger: Parakeet, by a wide margin.

Where you meet these models in apps

MacWhisper is built around Whisper and has added Parakeet as an option. Mac Note Taker uses Parakeet TDT v3 by default and offers Whisper Large v3, Turbo, Distil, and Small in Settings, so the same models are one dropdown apart. If you are choosing between those two apps, the engine is not the deciding factor; what happens around the transcript is.

The numbers above are what Mac Note Taker produces on a Mac it has tuned itself to, which is the figure worth comparing against.