Dictation speed on Apple silicon, measured

At a glance

Name	p50	p90	Takes	Languages
Parakeet v3	40 ms	47 ms	18	English
Qwen3-ASR 0.6B, with draft	30 ms	102 ms	12	Indonesian + English
Qwen3-ASR 0.6B, no draft	195 ms	302 ms	12	Indonesian + English
Whisper Turbo	529 ms	734 ms	12	Indonesian + English
Whisper Turbo	652 ms	698 ms	18	English (id,en set)

What we measured, and how

The number that matters in dictation is the wait after you stop talking: the time from releasing the key to having text ready to type. That is what most figures below measure. The engine is already loaded and warm, and the audio comes from real takes recorded by the developer, not synthetic test sets.

Most runs were on a MacBook Pro with an M5 Max in November 2026. Sample sizes are small, between 6 and 18 takes per test, so read them as a guide to scale rather than exact figures. They are Speakey's tests of its own engines. We haven't measured other apps and don't make claims about them.

p50 and p90, in plain words

p50 is the median: half the takes were faster than this, half slower. It tells you what a normal take feels like. p90 means 9 out of 10 takes were at least this fast. It tells you how bad the slow ones get. A good dictation engine has both numbers low and close together, because one slow take in ten is the one you notice.

Results by engine

Parakeet TDT 0.6B v3 on the Neural Engine finished 18 English takes at 40 ms p50 and 47 ms p90. Whisper Turbo on the same takes, set up for Indonesian and English, took 652 ms and 698 ms. Parakeet's sentences were as good or better, with punctuation, but it slipped on app and brand names and garbled every Indonesian take, because it only knows 25 European languages.

Qwen3-ASR 0.6B, in Speakey's own Core ML conversion, was tested on 12 mixed Indonesian and English takes. With a draft from its live preview, the final text took 30 ms p50 and 102 ms p90. Without a draft it took 195 ms and 302 ms. Whisper Turbo on those same 12 takes took 529 ms and 734 ms. The larger Qwen3-ASR 1.7B, on 6 takes, took 278 ms p50 plain and 152 ms with a draft, against 171 ms and 30 ms for the 0.6B, and made clearly fewer word errors.

Whisper Large v3 Turbo through WhisperKit on the Neural Engine averaged about 360 ms a take in one test and about 430 ms in another; the smaller Compact build took about 555 to 640 ms. Turbo's numbers vary between tests because the takes and the language settings differ: with several languages picked, Whisper also detects the language in the same pass.

Apple Speech with the SpeechTranscriber engine on macOS 26 took 57 to 147 ms across Speakey's tests, against 55 to 671 ms for Apple's older SFSpeechRecognizer.

Power: the Neural Engine versus the GPU

Whisper can also run on the GPU through whisper.cpp with Metal. In Speakey's test that was a little faster, about 250 ms a take against 360 ms for WhisperKit on the Neural Engine, with the same accuracy. But whisper.cpp pinned the GPU and drew about 74 W over idle, while WhisperKit drew about 28 W. On a laptop that difference is battery and fan noise, so Speakey uses the Neural Engine.

Live preview lag

While you hold the key, Speakey shows words as you say them. With Parakeet, it decodes the whole take so far every 200 ms. Measured on a real-time feed of 18 takes, words reached the screen 145 ms after being spoken on average, 255 ms at p90. A sliding-window approach using 0.5 to 2 second chunks lagged 291 to 1,281 ms and dropped 34 to 65% of words at release. Holding the key cost about 27% of the Neural Engine and about 1.2 W of CPU.

When you aren't talking

A dictation app sits in the background all day, so its idle cost matters more than its peak. With no key held and nothing on screen, Speakey measured 0.0% CPU and zero wakeups. An earlier build sat at about 40% because of animations running in hidden windows, which is the kind of thing worth checking in any menu bar app with Activity Monitor.

First run and older Macs

The first load of an engine compiles it for the Neural Engine: about 11 seconds for Parakeet, then about 170 ms per load, and about 60 seconds for Whisper, once per Mac. Qwen3-ASR loads in about 4 seconds.

Not every Mac is an M5 Max. On one user's base M-series Mac, Whisper Turbo took 2.8 to 3.9 seconds per Dictate take, and users on an M1 Pro or M2 saw about 1.8 to 2.1 seconds. That is why Speakey estimates Whisper's speed from your chip and measures it on first use, and suggests Apple Speech when Turbo would take more than 700 ms, or Parakeet or Qwen3-ASR when your languages allow.

Questions