Apple Speech vs Local Whisper on Mac: Accuracy and Latency Tested
We tested Apple Speech and a local WhisperKit base model on the same 30 labeled English clips. Here are the measured word error rate, first-text latency, finalization time, and the reason a Mac caption app may use both engines.
2026-08-30 · 9 minThe short answer
On this Apple-silicon Mac, Apple Speech produced its first text sooner, while local Whisper showed a different trade-off between model accuracy and the live caption pipeline. Across 30 clean English LibriSpeech clips, Apple Speech reached 0.9% WER, the Alison Workspace Whisper live pipeline reached 7.9%, and Whisper whole-clip recognition reached 1.9%. Median first-text latency was 1.14 s for Apple and 1.74 s for the Whisper live path. These numbers describe this test, not every speaker, language, Mac, or recording.
How the test was run
The test ran locally on August 30, 2026, on an Apple-silicon Mac with macOS 26.5.2. Both engines received the same 30 labeled clips from LibriSpeech test-clean, converted to 16 kHz mono audio. Frames were delivered at wall-clock speed in 100 ms pieces, followed by 0.8 seconds of identical silence so both engines saw the same endpoint. Models and Apple language resources were prepared before measured sessions. Accuracy used word error rate; latency measured the first non-empty text and the wait from the end of speech to final text.
Accuracy: separate the engine from the live pipeline
Whisper whole-clip recognition scored 1.9% WER and shows the model's accuracy ceiling when it can decode a complete utterance. The product's Whisper live path scored 7.9% because voice activity detection, segmentation, progressive drafts, and final assembly also affect what reaches the caption bar. Apple Speech's native streaming finals scored 0.9%. This distinction matters: a strong model can still lose words when a real-time pipeline cuts an utterance at the wrong place, and a single aggregate percentage cannot explain every error.
Latency: first text and final text are different
Apple Speech showed the first text after a median 1.14 s, compared with 1.74 s for the Whisper live path. After speech ended, Apple's native final arrived after a median 1.13 s, while Whisper's final arrived after 0.58 s. Whisper whole-clip decoding ran at a median real-time factor of 0.021×, but that offline-style speed is not a first-caption measurement. A useful live caption experience needs both an early readable draft and a dependable final line, so reporting only one latency number hides an important trade-off.
Why Alison Workspace uses both engines
Alison Workspace does not assume that one recognizer is best in every situation. On macOS 26, an explicitly selected language with an available Apple on-device speech resource can use Apple Speech's native streaming path. Older systems, automatic detection, unsupported Apple locales, or unavailable Apple resources can use the local WhisperKit path. This routing preserves broad recognition coverage while taking advantage of Apple streaming where it is available. After the required language resource or Whisper model is downloaded, recognition runs on the Mac rather than sending the session audio to a remote transcription service.
What this benchmark does not prove
The 30 clips are clean English read speech, mostly one speaker at a time. They do not represent every accent, meeting microphone, noisy room, music bed, overlapping conversation, specialist term, or any non-English language. Translating this article into other website languages does not turn the English test into a multilingual benchmark. We publish the sample count, operating system, model, input method, and separate pipeline results so readers can interpret the scope. For important work, test names, numbers, accents, and the exact app that plays your audio.
A one-minute test on your own Mac
Play a short video or meeting recording with one clear speaker. Select the known spoken language, confirm that the app receives Mac system audio rather than room microphone noise, and note when the first useful words appear. Check a name, a number, and a technical term, then compare the final line with the audio. Repeat with the meeting app, browser, or player you actually use. This small test is more informative than treating any benchmark as a guarantee, because your source audio and vocabulary determine the practical result.
Frequently Asked Questions
Is Apple Speech faster than Whisper on Mac?
Apple produced first text sooner in this 30-clip English test: median 1.14 s versus 1.74 s for the Alison Whisper live path. Other hardware, languages, audio, and pipeline settings can change the result.
Is local Whisper more accurate than Apple Speech?
Whisper whole-clip recognition reached 1.9% WER here, while Apple native streaming reached 0.9%. The Whisper live product path measured 7.9%, showing why model accuracy and end-to-end caption accuracy must be reported separately.
Does speech recognition work offline on Mac?
Yes, after the required Apple language resource or local Whisper model has been downloaded. Initial resource installation may require Internet access; recognition then runs on the Mac.
Which speech engine does Alison Workspace use?
It can use Apple Speech on macOS 26 when the selected language and on-device resource are supported, and it uses local WhisperKit for broader coverage or as a fallback.