Is Apple's new on-device speech model as accurate as Whisper?

No, not in our test. Apple's SpeechTranscriber made more word errors than Parakeet TDT v3 in every language where it produced text, and more than Whisper large-v3-turbo in all of them except French, where the two are tied. In English Apple had a word error rate of 6.73 percent ¹, against 3.92 percent for Whisper ² and 4.48 percent for Parakeet ³. In Portuguese Apple's output scored 100.00 percent ⁴, which means none of the words matched.

Apple's model is new: its SpeechAnalyzer interface, with SpeechTranscriber as the transcription module, was introduced at WWDC25 ⁵, and Apple says transcription with it is entirely on the device once the language models are downloaded ⁶. We measured it on the six languages it supports, next to the two open engines.

Apple is behind in the five languages where it worked

Word error rate on 100 recordings per language from the FLEURS test set, one Mac with an M2 Pro, percent, lower is better:

Apple's best results are Italian (4.15 percent) and Spanish (4.30 percent), which would be fine for dictation. But in the same two languages Parakeet and Whisper are lower still. The widest gap to Whisper is in English, where Apple's error rate is almost three points higher. In French Apple and Whisper are effectively tied (a difference of three hundredths of a point), and Parakeet is lower than both.

Apple's Portuguese run produced no usable text

Apple's Portuguese word error rate was 100.00 percent ⁴, and the run finished 51177 times faster than real time ²¹. That speed is not a real result: it means the engine returned almost immediately, with nothing useful. We treat it as a failed run, not as evidence that Apple's model cannot handle Portuguese. We did not investigate the cause. Whisper had 3.28 percent in Portuguese ²⁰ and Parakeet 3.62 percent ¹⁹.

The published median for Apple, 6.66 percent over its six figures ²², includes that failed run. Compare it with the other two medians only on the shared languages, as in the table, not against the 25-language medians of Parakeet (10.11 percent ²³) and Whisper (11.34 percent ²⁴).

All Apple rows:

Another benchmark found the opposite, on different audio

Not every test agrees. A 2026 blog report from the speech company Inscribe said Apple's SpeechAnalyzer beat every Whisper model it ships on the clean and noisy halves of LibriSpeech ²⁵. That test is English audiobook speech and compares Apple with smaller Whisper models. Ours uses a different recording set, a larger Whisper model and six languages. The two results can both be true: the winner depends on the audio and the model size.

What this means for dictation on a Mac

• If you dictate in English or German: Whisper or Parakeet gives fewer errors in our test. Apple's built-in model is still a reasonable free default, and it needs no extra app.
• If you dictate in Spanish or Italian: all three engines are good. Choose on convenience.
• If you dictate in Portuguese: do not assume Apple's model works. Test it on your machine first.
• If your language is not one of the six: Apple's model is not an option in our test, and the other two cover 25 languages. See which languages work best.

What the data does not show

• Read sentences in a quiet room. FLEURS is read speech. Real dictation has hesitations, names and noise.
• One Mac. All three engines ran on a Mac with an M2 Pro. Results on other chips and macOS versions may differ.
• Words only. These are plain word error rates. Counting punctuation and capital letters, Apple's English text differed from the reference by 11.10 errors per 100 tokens ²⁶, against 7.71 for Whisper ²⁷.
• A failed Portuguese run. One rerun on a clean setup could change the Portuguese row.

Related reading

Read the guide to on-device speech recognition and our overview of speech recognition. The Whisper against Parakeet comparison covers all 25 languages, is Apple Dictation better than Whisper looks at the built-in feature, and speed on a Mac compares the engines' processing time. Full rows are in the benchmark dataset, and speeds in the speed dataset.