Whisper vs Parakeet on Mac: measured on Apple Silicon
On a Mac with Apple Silicon, Parakeet Ultra transcribed about 7 times faster than Whisper Large v3 Turbo and was as accurate or more accurate in English, French and Spanish on dictation-length human speech. Whisper is still the pick for a language Parakeet does not cover, and Large v3 Turbo was the best Whisper model in this test.
The numbers come from one run on 25 Oct 2026, on an Apple M5 Pro with 48 GB of memory and macOS 27.0.1, using the exact engines and model files of YappYapp 0.4.1.
Results on human speech
The table uses human recordings cut into dictation-length clips (3.5 to 23 seconds, median 13), since that is what a dictation app transcribes. WER is word error rate, the share of words the model got wrong: lower is better.
| Model | Engine | Size on disk | Load time | Real-time factor | WER English | WER French | WER Spanish |
|---|---|---|---|---|---|---|---|
| Parakeet Ultra (YappYapp default) | FluidAudio 0.17.7, CoreML | 632 MB | 0.12 s | 0.0035 (284x) | 0.5 % | 1.9 % | 1.3 % |
| Parakeet TDT v3 | FluidAudio 0.17.7, CoreML | 491 MB | 0.11 s | 0.0034 (294x) | 0.5 % | 2.8 % | 2.6 % |
| Whisper Large v3 Turbo | whisper.cpp, Metal | 1,628 MB | 0.58 s | 0.0245 (41x) | 0.5 % | 1.9 % | 2.0 % |
| Whisper Small | whisper.cpp, Metal | 504 MB | 0.22 s | 0.0164 (61x) | 1.1 % | 13.0 % | 3.3 % |
| Whisper Base | whisper.cpp, Metal | 152 MB | 0.10 s | 0.0087 (115x) | 2.1 % | 21.3 % | 11.1 % |
| Qwen3-ASR 0.6B | llama.cpp, Metal | 1,025 MB | 0.52 s | 0.0224 (45x) | 0.5 % | 5.6 % | 5.9 % |
Real-time factor is transcription time divided by audio length: 0.0035 means one minute of audio takes about 0.2 seconds. For a single dictation-length clip, the median wait was 0.08 s with Parakeet Ultra and 0.60 s with Whisper Large v3 Turbo.
The human clips are public-domain audiobook readings, one speaker per language and 150 to 200 words each, so one word is about half a point. Differences of 1 or 2 points are not meaningful. The large gaps are Whisper Base and Small in French, and Base in Spanish.
What the numbers show
- Parakeet is the fast one: 7 to 8 times faster than Whisper Large v3 Turbo, on long and short clips.
- Turbo matches Parakeet on short clips, within noise in all three languages.
- Small Whisper models struggle in French. Base got about one word in five wrong, Small about one in eight.
- Qwen3-ASR is accurate in English and a few points behind in French and Spanish. It is in the app for other languages, which this test did not cover.
Whisper and recordings over 30 seconds
On the long clips (38 to 79 seconds), all three Whisper models, run with YappYapp's default flags, skipped or garbled whole sentences in this run, apparently around the 30-second boundaries of Whisper's decoding window. On the 204-word French human clip, Large v3 Turbo dropped 52 words, Small 94 and Base 73. Parakeet and Qwen3-ASR did not show this. Most dictations are under 30 seconds, which is why the table uses short clips, but if you dictate long passages, it is a reason to prefer Parakeet.
Memory
Whisper's peak memory was about 0.5 GB for Base, 1 GB for Small and 2 GB for Large v3 Turbo; Qwen3-ASR 0.6B peaked between 2 and 2.6 GB. There is no comparable Parakeet figure: CoreML runs its encoder on the Neural Engine, whose memory is held by macOS services outside the measured process, so the benchmark could not count it.
Synthetic voices flatter every model
Clips made with the macOS say voices were transcribed almost perfectly by every model, including Whisper Base (0 to 4.9 % WER). They cannot tell these models apart, so this page leads with human recordings. Real dictation, with hesitations, background noise and laptop microphones, will produce more errors than either kind of clip.
Which one to pick
- Parakeet Ultra for English, French, Spanish and most European languages. It is YappYapp's default and was the fastest model here, with accuracy level with the best. It covers 25 languages, including German, Italian, Portuguese, Dutch, Polish, Russian and Ukrainian, and detects the language itself. Parakeet TDT v3 has the same languages and speed; Ultra is a retrained v3 with fewer mistakes.
- Whisper Large v3 Turbo for a language Parakeet does not cover. Whisper's multilingual models handle about 100 languages. Avoid Base and Small for French, and Base for Spanish.
- Qwen3-ASR for Chinese, Japanese, Korean, Arabic, Hindi and similar. YappYapp 0.4.0 added it for those languages (52 languages and dialects), in 0.6B (about 1 GB) and 1.7B (about 2.5 GB) sizes. Tell setup you speak one of them and it proposes Qwen3-ASR.
You switch models on the MODELS tab; the setup guide shows where, and YappYapp is free on the download page. For dictating without a connection, read the offline dictation guide. Comparing apps rather than models? See YappYapp compared with Superwhisper.
How we measured
- Hardware: Apple M5 Pro, 48 GB, macOS 27.0.1, in normal use with other apps open, so timings are typical rather than lab-grade.
- Engines and models: what YappYapp 0.4.1 ships and pins. FluidAudio 0.17.7 for Parakeet (NVIDIA's model card), int8 encoder; whisper.
cpp commit 7695a53 with the app's flags (language auto-detect, beam size 5); llama.cpp v0.6.0 for Qwen3-ASR, with its prompt cache off so repeated runs could not reuse earlier work. Checksums verified. - Clips: six long clips (synthetic 38 to 39 s, human 72 to 79 s) for speed, and the same audio cut into 24 clips for accuracy and latency. Human English came from LibriSpeech, French and Spanish from Multilingual LibriSpeech.
- Timing: load time is the median of three launches with the model cached. The first load after installing a Parakeet model took about 12 to 13 s while macOS compiled it for the Neural Engine; Qwen's took 5.4 s. Speed is the median of three warm runs.
Not measured: Whisper Large v3, Whisper Medium, Qwen3-ASR 1.7B, Parakeet TDT v2 and Parakeet 110M. YappYapp's word list, cleanup and text insertion were not tested either; these are raw engine numbers.
FAQ
What is the best Whisper model for a Mac?
In this test, Whisper Large v3 Turbo: it matched Parakeet's accuracy on short clips at 41 times real time. Base and Small were faster but far less accurate in French.
Is Parakeet better than Whisper?
For the languages it covers, it was faster and at least as accurate here, and it did not drop text on long recordings. Whisper covers far more languages.
Do these models need the internet?
No. After the one-time model download, they all run on your Mac.