How accurate is speech recognition in English?

English speech recognition is accurate enough for dictation, and the engine you pick still changes the error rate by a wide margin. In our October 2026 benchmark on a Mac with an M2 Pro, Whisper large-v3-turbo had a word error rate of 3.92 percent ¹, Parakeet TDT 0.6B v3 had 4.48 percent ², and Apple SpeechTranscriber had 6.73 percent ³. A word error rate of 4 percent means about 4 wrong words in every 100. All three engines ran entirely on the device, on 100 English recordings from the FLEURS test set.

Whisper is the most accurate engine on the Mac, and Apple's is the fastest

The three engines trade accuracy for speed. Whisper large-v3-turbo was the most accurate in English but the slowest, at about 12 times faster than real time ⁴. Parakeet was close behind on accuracy and quicker, at about 19 times faster than real time ⁵. Apple SpeechTranscriber was the least accurate and the quickest, at about 34 times faster than real time ⁶.

For live dictation, where you speak a sentence and wait for the text, all three are faster than you talk. Speed matters more when you transcribe long recordings. Speeds for every engine and language are in the speed dataset.

Our measurement of all three engines in English

The table lists every engine we ran on English. Lower is better.

The full benchmark, with the transcripts behind each number, is on the on-device benchmark page.

Published results agree: English lands between 3 and 5 percent

Other groups test on the same FLEURS set, mostly with cloud models. The Voxtral paper reports 3.19 percent ⁷ for GPT-4o mini Transcribe, 3.29 percent ⁸ for ElevenLabs Scribe and 3.35 percent ⁹ for Voxtral Small. Whisper large-v3 scored 4.00 percent ¹⁰ in that paper, and NVIDIA reports 4.85 percent ¹¹ for Parakeet. Our on-device figures sit inside that range, so local engines give up little against cloud services on this kind of audio. Apple's built-in model is the exception.

Showing 8 of 319 rows. See every row in the dataset.

More rows, and how each was measured, are on the published accuracy page.

English is accurate, but not the most accurate language in the benchmark

You might expect English to top the ranking, since most speech models train on more English than anything else. It does not. Taking the best engine in each language, English comes fourth lowest of the 25 languages in the benchmark. Italian (2.52 percent with Parakeet ²⁰), Spanish (2.53 percent with Whisper ²¹) and Portuguese (3.28 percent with Whisper ²²) all score lower.

Part of the gap is the test itself. A scorer counts a spelling variant or a number written in digits as an error even when the engine understood the speech. Read the ranking as "English is in the accurate group", not as proof that English is easier to recognize.

What to expect when you dictate in English

With the best on-device engine in a quiet room, expect a few wrong words per 100, mostly names, product terms and numbers, which an engine has seen least. With Apple's built-in model expect noticeably more. For messages and notes that is a light edit. For long documents, choose the most accurate engine you can run. If you correct as you go and want instant text, a faster engine is a fair trade.

Our guide to European languages in speech to text explains how to choose, and the languages overview covers the rest.

What these numbers do not show

The recordings are people reading prepared sentences in a quiet room, scored on one Mac. Spontaneous speech, accents, background noise, overlapping talkers and technical vocabulary all raise error rates, and none of them are in this test. Each language has only 100 recordings, so a gap of a few tenths of a point between engines is not a reliable difference. A word error rate also weighs every word equally, though a wrong "not" costs more than a wrong "the".

For how to read these figures, see the guide on how accurate speech recognition is. To compare languages, read the same question for German and Spanish.