2026-07-11 · updated 2026-10-27 · Mat Silverstein
Whisper vs Parakeet v3 for local transcription
When I first wrote this in July, the answer was simple: Minutes used Whisper, and Parakeet stayed switched off. That has changed. Since version 0.31.0, Minutes runs NVIDIA's Parakeet v3 inside its own process on Apple Silicon Macs, and installing it takes one click. Whisper is still there, and for some people it is still the right engine. Here is how the two compare on published benchmarks and on our own audio, and where each one falls short.
What changed
The July version of this post was about parakeet.cpp, a local port of Parakeet that only accepts audio as a file path. Minutes never writes private audio to a named file that another process on your Mac could open, so we left that integration off. It is still off.
Parakeet v3 now runs through sherpa-onnx instead, loaded as a plugin inside the Minutes process. It receives audio the same way Whisper does, as samples in memory, so no new copy of your audio touches disk. The model is parakeet-tdt-0.6b-v3 in int8 form, a 639 MB download from a pinned Hugging Face revision. Minutes checks every file against its SHA-256 before using it.
The published numbers
NVIDIA released Parakeet v3 in August 2025. It has 600 million parameters, covers 25 European languages with automatic language detection, writes punctuation and capitals, and is licensed CC BY 4.0. NVIDIA's paper reports these word error rates on the Hugging Face Open ASR Leaderboard (lower is better):
| Model | Average | AMI meetings | LibriSpeech other | Speed (RTFx) |
|---|---|---|---|---|
| Whisper large-v3 | 7.44% | 15.95% | 3.91% | 146 |
| Parakeet v3 | 6.32% | 11.39% | 3.59% | 3,333 |
RTFx is how many seconds of audio a model gets through per second of compute on NVIDIA's test GPU. The AMI column is the one I care about most, because it is recorded meetings. There Parakeet v3 makes about 29% fewer errors than Whisper large-v3, and it does it with well under half the parameters.
Our numbers, on a Mac
Minutes defaults to Whisper small, not large-v3, so we also measured the models Minutes users actually run. This was on an M4 Max with the models already loaded, using 39 clips from LibriSpeech test-other:
| Engine | Word error rate | Median time per clip |
|---|---|---|
| Parakeet v3 (int8) | 3.46% | 151 ms |
| Whisper medium | 4.45% | 759 ms |
| Whisper small (Minutes default) | 5.81% | 290 ms |
LibriSpeech is people reading books aloud, with few names or acronyms, so I also ran 39 of my own saved dictation clips. There is no hand-checked transcript for those, so Whisper medium served as the reference. Parakeet v3 differed from it on 2.1% of words and Whisper small on 6.9%. Parakeet decoded a typical clip in 51 ms, Whisper small in 168 ms.
Where Whisper still wins
Languages. Whisper covers 99 languages and Parakeet v3 covers 25, all European. If you set a language Parakeet does not cover, such as Chinese, Japanese, Hindi or Arabic, Minutes keeps Whisper and never offers the install.
Platforms. The desktop app uses Parakeet only on Apple Silicon Macs. On Linux, the command-line sherpa archive runs it if you set the engine to sherpa yourself. Windows and Intel Macs use Whisper for now; Windows support is tracked in #645.
Unusual names. Parakeet v3 can misspell rare names. In my dictation clips it wrote "Garek" for "Garrett", a name Whisper small got right. Minutes can steer Whisper toward the names in your vocabulary list, and Parakeet has no equivalent. What works for both engines is listing the misspellings you see: Minutes corrects them in every new transcript.
Live text. While you dictate, the words appearing on screen still come from Whisper; Parakeet writes the final text when you stop. Standalone live transcript runs on Whisper, or on Apple Speech if you opt in on macOS 26.
What Minutes picks by default
The transcription engine defaults to Auto. Auto uses Parakeet v3 when it is installed, you are on Apple Silicon, and your language is one Parakeet covers. Otherwise it uses Whisper. If Parakeet fails on a recording, Whisper transcribes it instead, and since 0.31.0 that includes the rare recording that crashes the Parakeet engine outright.
You can install Parakeet v3 from the banner Minutes shows when it is available, from Settings › Transcription, or during first-run setup. From the command line it is minutes setup --sherpa.
The recommendation
If you are on an Apple Silicon Mac and your meetings are mostly in one of Parakeet's 25 languages, install it. On our tests it was more accurate than Whisper small, the default, and about twice as fast. If your meetings are full of unusual names, add them to your vocabulary along with the misspellings you see. Everyone else should stay on Whisper, and with the engine on Auto, Minutes makes that call for you.
Setup details are in docs/architecture/sherpa-engine.md, and the parked parakeet.cpp integration is documented in docs/architecture/parakeet.md. The leaderboard figures come from Table 5 of NVIDIA's paper; the Mac measurements are ours. Minutes is MIT licensed, and the pipeline code is on GitHub.