Not as accurate as the number on the box. Both leading on-device speech models publish a single headline word error rate — 6.34% for NVIDIA's Parakeet, 7.44% for OpenAI's Whisper large-v3 — but those are averages across eight datasets dominated by read audiobooks. On the one dataset made of actual meetings, the same models score 11.31% and 15.95%.
That gap is the whole article. If you are choosing a Mac notetaker, or deciding whether on-device transcription is good enough for client calls, the meeting number is the one that describes your Tuesday.
Why isn't the WER you're quoted the one you'll get?
Because the standard benchmark averages meetings together with audiobooks. The Open ASR Leaderboard — the source of nearly every "X% WER" claim in this category — composites eight English datasets: LibriSpeech clean and other (read audiobooks), TED-LIUM (rehearsed talks), VoxPopuli (parliamentary speech), SPGISpeech and Earnings-22 (financial calls), GigaSpeech (mixed web audio), and AMI (meetings).
Only one of those eight is spontaneous multi-party conversation. The average is therefore a blend in which read and oratorical speech outnumber meeting audio seven to one, and read speech is the easiest audio there is.
Look at what that does to a single model. These are NVIDIA's own published figures for Parakeet TDT 0.6B v3, from its model card:
| Dataset (Parakeet TDT 0.6B v3) | What it is | WER |
|---|---|---|
| LibriSpeech test-clean | Read audiobooks | 1.93% |
| LibriSpeech test-other | Noisier audiobooks | 3.59% |
| GigaSpeech | Podcasts, YouTube, audiobooks | 9.59% |
| AMI | Real meetings | 11.31% |
| Earnings-22 | Earnings calls | 11.42% |
| Headline average | All eight | 6.34% |
The model is 5.9 times worse on meetings than on the read speech that anchors its headline number. Nothing is wrong with the benchmark — it is doing what it says. The problem is that marketing pages quote the composite and readers hear "accuracy on my calls."
How much worse is meeting audio, exactly?
By the vendors' own numbers, roughly two to eight times worse depending on what you compare against. Here is the same cut for both engines a Mac notetaker realistically ships, each taken from the model's own card rather than a third party:
| Parakeet TDT 0.6B v3 | Whisper large-v3 | |
|---|---|---|
| Headline average WER | 6.34% | 7.44% |
| AMI (meetings) WER | 11.31% | 15.95% |
| Meetings vs own average | 1.8× worse | 2.1× worse |
| Languages | 25 European | 99 |
| Licence | CC-BY-4.0 | MIT |
On English meeting audio Parakeet is meaningfully ahead — 11.31% against 15.95% is the difference between roughly one error in nine words and one in six. That ordering is the reverse of what many people assume from Whisper's larger size (1.55B parameters against 600M).
Two honest caveats about even these numbers. First, the AMI figures come from the corpus's headset-microphone condition, where every speaker wears their own mic — a more favourable setup than one laptop microphone in a room. Second, the AMI corpus is recorded with "mostly non-native speakers" across three rooms with different acoustics, which pushes the other way.
What actually makes a meeting hard?
Spontaneity and overlap. The AMI Meeting Corpus is 100 hours of real meetings released under CC BY 4.0, captured with synchronised "close-talking and far-field microphones" — built precisely because scripted audio was not telling researchers anything useful about conversation.
What a model hits in a meeting that it never hits in an audiobook: two people starting a sentence at once, half-words and restarts, company-specific proper nouns absent from training data, and the filler speech that makes transcripts look messy even when every word is right.
There is also a failure mode specific to Whisper's architecture, which OpenAI documents plainly on its own model card:
the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination)
OpenAI attributes this to the model "trying to predict the next word in audio" as well as transcribe it, and notes it is "prone to generating repetitive texts." In practice this surfaces during silence and crosstalk — exactly what a meeting has plenty of. It is why Speechmark filters hallucination-shaped segments out of Whisper output before a transcript is ever written to disk, rather than trusting the raw decode.
NVIDIA is similarly direct about Parakeet's limits: "Accuracy varies based on language and characteristics of input audio (Domain, Use Case, Accent, Noise, Speech Type, Context of speech, etc.)."
Where Whisper is the better choice
Languages, mainly — and it is not close. Whisper large-v3 covers 99 languages; Parakeet TDT 0.6B v3 covers 25 European ones, and Hindi, Mandarin, Japanese, Arabic, and Korean are not among them. If your meetings are not in that European set, the accuracy comparison above is irrelevant and Whisper is the only one of the two that works at all.
Whisper is also the more forgiving model on heavy accents and code-switching, where Parakeet's speed advantage stops mattering because the output needs correcting anyway. And OpenAI publishes per-language figures — 26.8% WER on Hindi, for instance — so you can check your own case rather than guess.
Parakeet wins on English meeting accuracy and on speed. Whisper wins on coverage and on graceful degradation. That is why Speechmark ships both as a user-selectable engine instead of picking one, and routes to Whisper automatically for languages Parakeet cannot handle.
What does this mean for a Mac notetaker?
That architecture matters more than the leaderboard. A notetaker that captures each source separately gets closer to the benchmark's favourable headset condition than one recording a room.
Speechmark records your microphone and the call's system audio as two independent streams via a Core Audio process tap, then transcribes each one on its own before merging them into a single speaker-attributed transcript. For a remote call that means the far end arrives as a clean digital feed rather than as room audio picked up by your laptop — structurally closer to AMI's per-speaker headset condition than to a single distant microphone.
That is a design choice, not a measured claim, and worth stating precisely: the figures in this article are the model vendors' published benchmark results, not Speechmark's own measurements. We have not run these models against our own labelled meeting set, and we would rather say so than dress up someone else's numbers as original research.
The practical read, as of August 2026: expect a good on-device meeting transcript to be excellent for search, summaries, and recall, and to need a human eye before you treat it as a verbatim record. That is true of cloud transcription too — the difference is where the audio goes to get there, and who keeps a copy.
Speechmark transcribes meetings entirely on your Mac with either engine, keeps the audio on your machine, and costs one payment instead of a subscription. See how it compares.