Which speech recognition handles accents best?

Parakeet TDT v3 handled accents best in our October 2026 test. Pooled over all 13 accent groups, it had a word error rate of 12.57 percent ¹. Apple SpeechTranscriber had 16.39 percent ² and Whisper large-v3-turbo had 17.66 percent ³. Lower is better. All three engines ran on the same Mac with an M2 Pro, on the same recordings.

Parakeet leads when all 13 groups are pooled

The pooled figure covers 47 speakers and 1,224 conversational utterances from the EdAcc corpus (English video calls between speakers of many accents, University of Edinburgh, 2023). Every engine transcribed exactly the same utterances.

Parakeet's lead over the other two is larger than the gap between those two. Apple is ahead of Whisper, but only by a small margin next to Parakeet's advantage. Because the Apple and Whisper figures sit close together, the order of those two depends on which accents you care about (see below).

Parakeet is the lowest in every single group

The pooled number could hide a split, so we checked each group. There is none. Parakeet had the lowest word error rate in all 13 groups. Examples:

• Irish English: Parakeet 9.35 percent ⁴, Whisper 11.24 percent ⁵, Apple 11.49 percent ⁶.
• Mainstream US English: Parakeet 12.32 percent ⁷, Apple 15.45 percent ⁸, Whisper 16.93 percent ⁹.
• Vietnamese speakers: Parakeet 22.76 percent ¹⁰, Apple 25.71 percent ¹¹, Whisper 28.21 percent ¹².

The full picture, group by group, is in the chart below.

Apple and Whisper trade second place

Apple had the second-lowest rate in nine groups and Whisper in four. Whisper was ahead of Apple in Southern British English, Scottish English, Irish English and Chinese speakers. Apple was ahead in the other nine, including Mainstream US English, Indian English, Nigerian English and Vietnamese speakers. In all four of the Whisper groups the two engines were separated by less than a point, so treat that order as a tie.

Within one engine, the spread across accent groups is wider than the pooled gap between engines. Parakeet ran from 8.31 percent for Latin American Spanish speakers ¹³ to 22.76 percent for Vietnamese speakers ¹⁰. Who is speaking matters at least as much as which engine you pick.

What the numbers do and do not measure

• Few speakers per group. Groups rest on between 2 and 9 speakers. One person's voice, microphone and topic weigh heavily, so compare engines within a group more than groups with each other. The ranking of engines is the sturdier result because every engine heard the same people.
• Conversation, not dictation. The recordings are casual talk between friends over video calls. Overlapping speech was removed, but the talk is unscripted and informal, so these rates are not a forecast for your own dictation.
• Digits count as errors. Numbers spoken as words in the reference count as errors when an engine writes digits. This affects all engines, but not equally.
• Selected utterances. We used at least five words, 1.5 to 30 seconds, no overlapping speech, and up to 100 per group. Accent groups with a single speaker are left out, and so are cloud services.
• One test, one machine. The results describe these three engines on a Mac in October 2026. Newer versions can change them.

What this means for choosing a tool

If accuracy on varied accents is your main concern and you can run Parakeet, it is the safest pick on this evidence. If you use the free built-in option on a Mac, Apple SpeechTranscriber was second overall, and for some varieties of British, Scottish and Irish English Whisper was marginally lower. The more useful habit is to test an engine on your own voice for a few minutes, since the accent spread is large. Our own app, Yap, is on-device dictation for Mac and Windows; see /market for how products compare.

Related reading: the full dataset, the noise robustness data, the on-device benchmark, dictation across languages and the guide to speech recognition and accents.