Speech Recognition and Accents: 13 English Accents Tested

An accent can change how many words an on-device engine gets wrong by a wide margin, and which engine you use changes it by several points in most groups. Across all 13 accent groups, Parakeet TDT v3 had the lowest error rate at 12.57 percent ¹. Apple SpeechTranscriber followed at 16.39 percent ², and Whisper large-v3-turbo had the highest at 17.66 percent ³. The data holds 42 data points ⁴: 14 readings for each of the three engines, one per accent group plus one pooled reading.

All three were run on a Mac with an M2 Pro in October 2026 on identical utterances from the EdAcc corpus, English conversations over video calls. The measure is word error rate, the share of words wrong, so lower is better.

Parakeet had the lowest error in every one of the 13 groups

Parakeet's lead is not an average that hides exceptions. In each accent group its error rate was below both other engines. Whisper and Apple are harder to rank. Apple was lower than Whisper in 9 of 13 groups, and Whisper was lower in 4: Southern British English (17.53 against 17.91 percent), Scottish English (15.49 against 15.65 percent), Irish English (11.24 against 11.49 percent) and Chinese speakers (13.50 against 14.11 percent). In those four the margins are under a point. In the groups where Apple was ahead, the margins are often larger, so the pooled gap between the two is smaller than the gap you see in some single groups.

Vietnamese speakers had the highest error rate with all three engines

Every engine did worst on the same group. For Vietnamese speakers, Parakeet reached 22.76 percent ⁵, Apple 25.71 percent ⁶ and Whisper 28.21 percent ⁷. For each engine, no other group came close to that figure. This group also has 6 speakers, more than any other except Mainstream US English with 9, so it is less likely than most to reflect one unusual voice.

The lowest readings differ by engine. Latin American Spanish speakers were lowest for Parakeet (8.31 percent ⁸) and Apple (10.77 percent ⁹), while Irish English was lowest for Whisper (11.24 percent ¹⁰). The Latin American group rests on 52 utterances from 2 speakers, fewer than the up to 100 used elsewhere, so treat it as the least firm of the best readings.

Being a native speaker of English did not guarantee the best result

A natural guess is that people whose first language is English are transcribed best and everyone else worse. The rows do not show that. Mainstream US English, the group with the most speakers, was not the easiest for any engine. Parakeet read it at 12.32 percent ¹¹, above its readings for Irish English (9.35 percent), Indian English (9.82 percent), Chinese speakers (9.70 percent) and Lithuanian speakers (10.54 percent). Whisper shows the same pattern: Mainstream US English at 16.93 percent ¹² sat above Chinese speakers at 13.50 percent and Indian English at 14.26 percent ¹³.

The pooled reading for each engine is also slightly higher than its Mainstream US reading, in each case by less than a point, so the typical accent in this test is a little harder than the US baseline, not a lot.

The kind of accent, a variety of English or a first language other than English, does not sort the groups either. Both kinds appear among the lowest readings and both among the highest. The two Spanish groups show how much a group's makeup matters. Spanish speakers (5 speakers) had 13.93 percent with Parakeet ¹⁴, 17.01 percent with Apple ¹⁵ and 18.82 percent with Whisper ¹⁶. Latin American Spanish speakers, also with Spanish as a first language, scored lower with all three engines. Whatever separates those two groups, it is not "Spanish as a first language" alone: it is also the particular people, microphones and topics.

The engines disagree most on Nigerian English and Lithuanian speakers

The engines do not rank accents the same way, and the gaps between them vary by group.

• Nigerian English: Whisper 22.03 percent ¹⁷, Apple 20.20 percent ¹⁸, Parakeet 13.44 percent ¹⁹. Parakeet's result here is lower than the Mainstream US result of Apple (15.45 percent ²⁰) and of Whisper.
• Lithuanian speakers: Whisper 19.33 percent ²¹, Apple 13.81 percent ²², Parakeet 10.54 percent ²³. This is the widest gap between Whisper and Parakeet in any group.
• Southern British English: Parakeet 11.83 percent ²⁴ was lower than its Mainstream US reading, but Whisper (17.53 percent) and Apple (17.91 percent) were both higher than their Mainstream US readings. So whether British English is easier than US English depends on the engine.

Measured from best group to worst group, Whisper has the widest range of the three and Parakeet the narrowest, in percentage points. Parakeet is not only the lowest overall but also the least sensitive to who is speaking. That sensitivity is an engine property you can check before you rely on it.

What these numbers cannot tell you

• Few speakers per group. The pooled reading rests on 47 speakers. Many groups have only 2 or 3, and the Mainstream US group, the largest, has 9. One speaker's voice, microphone and topic weigh heavily, so compare engines within a group more than groups with each other. A gap of a point or so between two groups is not a finding.
• Only groups with at least two speakers. Accent groups with a single speaker in the test set, such as Ghanaian English, Israeli, Indonesian, Italian, Romanian, French and Catalan, are left out. The results say nothing about them.
• Digits count as errors. Numbers spoken as words in the reference count as errors when an engine writes digits. Every reading is inflated by this for every engine, and an engine that formats numbers differently can lose points it would not lose in everyday use.
• Short, clean conversational clips. Utterances are 1.5 to 30 seconds, at least five words, without overlapping speech, from video calls. Long dictation, background noise and technical vocabulary are not measured here. For noise, see our data on speech recognition in noise.
• Local engines only. Cloud services are excluded, so this does not tell you how a server-based service would do on the same speech.
• One test, one date. One machine and one month. Engines are updated, so a result can age.

What it means if you dictate in an accent

Choose by testing, and expect the engine to matter. If your accent is one of the harder ones here, the best step is to compare engines on your own voice, not to assume a ranking from a different group. In this test, Parakeet's pooled error of 12.57 percent was well below Whisper's 17.66 percent, and Apple sat in between. Apple's pooled error is only a little lower than Whisper's, and in four groups Whisper came out lower, so either can be the better pick for you.

A few habits help whichever engine you use:

1. Record a short passage of your usual speech, the kind you would actually dictate, and read the result with a red pen. Count only the errors that matter to your text.
2. Check digits, dates and names after dictating. Engines write numbers in different ways, and that alone shifts error counts.
3. Expect noise to add errors on top of the accent. Our noise robustness data shows how much.
4. For a wider view of engine speed and accuracy on standard test sets, see the on-device benchmark. For dictating in a language other than English, see our overview of languages.

The accent effect is real, but it is smaller than the spread between a good and a poor choice of engine for the hardest groups. In this test, Parakeet on Nigerian English (13.44 percent) did better than Whisper on Mainstream US English (16.93 percent). Choosing the right engine matters at least as much as the speaker's accent does.

Every reading

The table lists each engine for every accent group, with the number of speakers and utterances behind it.