How much faster is Parakeet than Whisper?

In our benchmark on Apple Silicon, Parakeet was several times faster than whisper.cpp (compact model) on the CPU, and about four times faster than Whisper with Metal acceleration. In that CPU test Whisper ran with 4 threads and Parakeet with 8, so it is not a thread-for-thread comparison. The exact numbers and the method are on the engine benchmark page; on other hardware the gap will differ.

Which languages does each engine support?

Whisper supports 99 languages. Parakeet covers around 25 European languages. If your language isn't on Parakeet's list, Whisper is the one to use.

Can I switch engines without losing my settings?

Yes. Both engines live under the provider "Locally on Device" in Settings → Speech Recognition. Switching takes one click, and your hotkeys, snippets and history stay exactly as they are.

Parakeet vs Whisper: Which Is Better for Local Dictation?

Whisper: the language all-rounder

Whisper runs as whisper.cpp, compiled directly into WhisPaste via dart:ffi: no subprocess, no server. It covers 99 languages in three quality tiers (compact, balanced, premium), so you can trade download size and speed for accuracy. On supported hardware it accelerates via Vulkan on Windows and Linux or Metal on Apple Silicon, and falls back to CPU automatically if a GPU crashes or is unavailable.

Parakeet: fast on GPU and CPU

Parakeet runs NVIDIA's Parakeet-TDT model, also fully inside the app. With a GPU it uses the whisper.cpp backends (Metal on Mac, Vulkan on Windows and Linux, CUDA with the optional CUDA pack); without a usable GPU it runs on the CPU through sherpa-onnx. It covers around 25 European languages (fewer than Whisper). In our Apple M5 CPU benchmark it was several times faster than the compact whisper.cpp model on the same CPU (Whisper ran with 4 CPU threads, Parakeet with 8), and even faster than Whisper with Metal acceleration. The download is 639 MB for the CPU bundle; with GPU acceleration enabled, the GPU model adds about 638 MB.

See the measured numbers on the engine benchmark page

Which engine should you use?

Reach for Whisper when you need a language outside Parakeet's list, or want to choose between model sizes. Reach for Parakeet when you mainly speak a supported European language and want the fastest result, especially without a capable GPU. On clear English read speech in our 20-utterance test, Parakeet made about half as many word errors (1.9% vs 3.4% WER). You can switch any time in Settings → Speech Recognition, with no reinstall and no data loss.

How much faster is Parakeet than Whisper?: In our benchmark on Apple Silicon, Parakeet was several times faster than whisper.cpp (compact model) on the CPU, and about four times faster than Whisper with Metal acceleration. In that CPU test Whisper ran with 4 threads and Parakeet with 8, so it is not a thread-for-thread comparison. The exact numbers and the method are on the engine benchmark page; on other hardware the gap will differ.
Which languages does each engine support?: Whisper supports 99 languages. Parakeet covers around 25 European languages. If your language isn't on Parakeet's list, Whisper is the one to use.
Can I switch engines without losing my settings?: Yes. Both engines live under the provider "Locally on Device" in Settings → Speech Recognition. Switching takes one click, and your hotkeys, snippets and history stay exactly as they are.