Is speech recognition worse for non-native English speakers?

No, not as a blanket rule. In our October 2026 test, the groups whose first language is not English did not all fare worse than the varieties of English. One group stood out as hard: Vietnamese speakers. Parakeet TDT v3 had a word error rate of 22.76 percent on them ¹, Apple SpeechTranscriber 25.71 percent ² and Whisper large-v3-turbo 28.21 percent ³. Each of those is the highest figure of any group for that engine. Lower is better.

Vietnamese speakers were hardest for all three engines

Vietnamese is the largest non-native group in the test, with 6 speakers and 100 utterances. For Parakeet, its rate is above every variety of English we measured (the highest of those is Jamaican English at 16.01 percent ⁴). For Apple, it is above the highest variety, Nigerian English at 20.20 percent ⁵. For Whisper, it is also above the highest variety, again Nigerian English at 22.03 percent ⁶.

Other non-native groups were not worse than Mainstream US English

Latin American Spanish speakers and Chinese speakers had lower error rates than Mainstream US English with all three engines. With Parakeet, the figures were 8.31 percent ⁸ and 9.70 percent ⁹, against 12.32 percent for Mainstream US English ¹¹. With Apple, Latin American Spanish speakers had 10.77 percent ¹², against 15.45 percent for US English ¹³.

Spanish speakers (5 speakers) went the other way. They were above Mainstream US English for every engine, for example 13.93 percent with Parakeet ⁷ and 18.82 percent with Whisper ¹⁴.

So the label "first language other than English" does not predict the outcome. The results split into low, middle and high groups, and the label alone does not tell you which one a speaker lands in.

The engine changes the answer for some groups

Lithuanian speakers (2 speakers) show it best. Parakeet had 10.54 percent ¹⁰ and Apple 13.81 percent ¹⁵, both below their Mainstream US English figures. Whisper had 19.33 percent ¹⁶, above its Mainstream US English figure of 16.93 percent ¹⁷. The same speakers were easy for two engines and harder for the third.

What the numbers do and do not measure

• Very few speakers. The five non-native groups are built from 2 speakers (Latin American Spanish, Chinese, Lithuanian), 5 speakers (Spanish) and 6 speakers (Vietnamese). One person's voice, microphone and topic can move a group's figure a lot. Compare engines within a group more than groups with each other.
• No pooled non-native figure. We report each group separately. The data does not support one number for "non-native speakers" because the groups differ so much.
• The label is the dataset's. The split follows each group's label in the EdAcc corpus, not an assessment of each person's English.
• Conversation over video calls. The recordings are informal talk, selected for at least five words, 1.5 to 30 seconds and no overlapping speech. They are not dictation in a quiet room.
• Digits count as errors. Numbers spoken as words in the reference count as errors when an engine writes digits.
• Left out. Accent groups with a single speaker and cloud services are not in this test.

What this means if English is your second language

Do not assume your accent will be a problem, and do not assume it will not. Your result may differ sharply from a group average. Try your engine on a few minutes of your own speech, and compare at least two engines if accuracy matters, since for Lithuanian speakers the engines were far apart. If you must pick without testing, Parakeet had the lowest error rate in every non-native group here; see which speech recognition handles accents best.

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