Why AI Transcription Keeps Getting Your Accent Wrong — And What Actually Fixes It

There is a specific kind of quiet frustration that shows up in meeting transcripts. Your colleague from Ohio gets a clean, punctuated paragraph. You — the one with the Bavarian, Taiwanese, Nigerian, Scottish, or Paulista accent — get a paragraph that reads like it was assembled by someone who heard the meeting through a wall. Your name is spelled three different ways. The number "thirteen" became "thirty". A whole clause is simply missing.

It's tempting to read that as a personal failure of pronunciation. It isn't. It's a measurable, documented property of nearly every automatic speech recognition (ASR) system shipping today, and the size of the gap is much larger than most people assume.

The gap, in numbers

Speech recognition quality is usually reported as word error rate (WER): the percentage of words the system inserts, deletes, or substitutes relative to a human reference transcript. A 5% WER means one word in twenty is wrong — annoying but readable. A 30% WER means roughly one word in three is wrong, which is where a transcript stops being a record and starts being a puzzle.

A 2025 evaluation of ASR on non-native English (arXiv:2503.06924) put concrete numbers on the split. On carefully read speech, native speakers came in at 8.0% WER while non-native speakers hit 24.8% — roughly three times the errors on identical material. Move from read sentences to actual dialogue, with its hesitations, restarts and overlapping turns, and both groups get worse: about 30.7% and 33.8% respectively. Practitioner benchmarks tell the same story from the other end: clean native speech on a good microphone lands in the 2-8% range, while strongly accented speech in a real room routinely lands in the 30-50% range.

The bias exists even within native English. A peer-reviewed evaluation of OpenAI's Whisper published in JASA Express Letters (2024) found systematically better recognition of American English than of British or Australian English, and found that accuracy scaled with how much lifetime English exposure a speaker had. Whisper is the model family behind a large share of the transcription tools you use, including on-device ones — so this is not a story about one bad vendor.

The stakes go past inconvenience. A 2026 study in npj Digital Medicine looked specifically at accent-related errors in clinical speech transcription and found error rates significantly higher for non-native speakers — in a setting where a substituted word can be a dosage.

Why models are worse at your voice

Four mechanisms explain most of the gap, and understanding them tells you which ones you can actually work around.

1. Training-data imbalance. Large ASR models learn from whatever audio the internet offers in bulk: American podcasts, YouTube, US broadcast media. The distribution of accents in that pile does not match the distribution of accents among the people in your meetings. The model isn't biased because someone chose bias; it's biased because it heard one accent a hundred times more often.

2. Phoneme mapping from your first language. Every speaker maps unfamiliar sounds onto the nearest sound in their native inventory. German speakers devoice final consonants, so bad drifts toward bat. Mandarin and Cantonese speakers often reduce final consonant clusters, so tests and test collapse. Brazilian Portuguese speakers frequently add a vowel after a final stop — hope becomes closer to hopie. Each of these is perfectly intelligible to a human listener with two seconds of adaptation. The model has no such adaptation step.

3. Prosody and tonal interference. Speakers of tonal languages carry pitch contours into English that the model reads as sentence boundaries or emphasis. Evaluations have repeatedly found some of the highest error rates for speakers whose first language is tonal. It's also why punctuation in your transcript can be strangely wrong even when the words are right.

4. The disfluency tax. Non-native speakers, quite reasonably, produce more mid-sentence repairs — restarts, filler, self-corrections — while composing in a second language. Models trained mostly on polished speech handle repairs badly, often deleting the corrected phrase along with the repair. This is why the errors cluster in your longest, most important sentences rather than in your short ones.

What actually breaks in a meeting transcript

Aggregate WER hides where the damage lands. In real meeting transcripts the errors are not evenly spread — they concentrate exactly on the tokens you'd want to keep:

• Proper nouns. Names of people, clients, products and cities are low-frequency tokens with no context to fall back on. A single meeting can spell one colleague's name four ways.
• Numbers and units. Thirteen/thirty, fifteen/fifty and fifty/fifty-five differ by a stress pattern that accent shifts. In a budget or dosage discussion, this is the single most dangerous error class.
• Domain jargon. Internal acronyms and product codenames were never in the training data at all, so the model substitutes the nearest common English word.
• Code-switching. Drop a German or Portuguese phrase into an English sentence and most systems, locked to one language for the whole file, will render it as English-sounding nonsense.
• Turn boundaries. When two accented speakers overlap, diarization frequently merges them, and a sentence gets attributed to the wrong person — which is worse than losing it.

The fixes that actually move the number

You can't retrain the model, but the research is clear that the gap is partly acoustic and partly configuration — and both of those are yours to control. Ranked roughly by effect size:

1. Shorten the mic distance. This dominates everything else. A phone lying flat in the middle of a six-person table is capturing reverberation as much as speech; the same phone propped 30-50 cm from the speaker, screen up, on something soft rather than a hard resonant table, can cut error rates dramatically. Accent bias and bad acoustics compound — fixing the second buys back headroom for the first.
2. Set the language explicitly. Auto-detection is a guess made on the first few seconds of audio, and a strongly accented opening sentence is exactly the case it gets wrong. If your tool lets you pin the language, pin it. If your meeting is genuinely bilingual, pin the language that carries the substance.
3. Feed it a vocabulary list. Any tool that accepts custom terms, a glossary, or a prompt with participant names will use them. Names and jargon are the highest-value errors to eliminate and the easiest, because they're a finite list you already know before the meeting starts.
4. Say numbers twice, once decomposed. "Thirteen — one three." It feels pedantic for exactly one meeting, and then it becomes the thing that saved you from a wrong figure in a summary.
5. Keep the audio, not just the transcript. The single most useful property of any transcription setup is the ability to jump back to the original audio at the timestamp where the text looks wrong. A tool that discards the recording after processing has removed your only correction mechanism.
6. Let an LLM clean up afterwards. The npj Digital Medicine work found that a language-model pass over an accented transcript recovers a meaningful share of accent-driven errors, because context resolves what acoustics couldn't. This is what a good summarization step is quietly doing for you already.

How to test a transcription tool on your own voice in ten minutes

Published benchmarks are averages over speakers who are not you. Vendor accuracy claims are almost always measured on clean native-accented read speech, which is the easiest possible condition. Run your own test instead — it takes one coffee break and it's the only number that matters:

1. Write a 150-word script that contains five colleagues' names, five internal acronyms, and six numbers including a thirteen/thirty pair.
2. Record it three times with the same tool: once at 30 cm in a quiet room, once at two metres in the same room, once with normal office background noise.
3. Count only substitutions and deletions on the names, acronyms and numbers. Ignore filler words — nobody cares about a missing "um".
4. Repeat with the second tool on your shortlist, using the same audio conditions.
5. Then check the recovery path: can you play back the audio at a wrong word, and can you fix the text and keep it?

The three-distance sweep usually surprises people more than the tool comparison does. If your 30 cm result is fine and your two-metre result is unusable, the problem was never your accent — it was the table. And if you're weighing specific products, our breakdown of Otter vs. Rev vs. Meetly covers how the cloud and on-device options differ on exactly this recovery path.

The multilingual meeting problem

The hardest case isn't an accent at all — it's a meeting that switches languages mid-sentence, which is the normal state of affairs in Zurich, Singapore, São Paulo and half of the internet's engineering teams. Almost every ASR pipeline decides on one language per file. Once that decision is made, the German clause in your English sentence is transcribed as English-shaped gibberish, and no amount of microphone discipline recovers it.

There are only two practical strategies. Either split the recording — stop and start a new file when the language changes, which is clumsy but works perfectly — or accept a transcript that is a memory aid rather than a record, and rely on a summary generated over it to carry the meaning across. Choosing the second is entirely reasonable, as long as you make the choice deliberately instead of discovering it a week later when you reread the file.

Why this is worth solving rather than tolerating

Memory research is unambiguous about the alternative. Without an external record, the majority of the specifics of a conversation are gone within a day — the reason we wrote a whole piece on why you forget meetings. If you're operating in a second language, you're carrying an extra cognitive load during the meeting and a worse transcript afterwards. The two failures compound, and the person they compound on is the one who already had the harder job in the room.

That's the real argument for taking mic distance, explicit language settings and a name glossary seriously. These aren't audiophile fussiness; they are the difference between a record you can rely on and one you quietly stop opening. If you'd rather start from a setup that already gets the microphone and the language settings right, we've written up what to look for in a private AI meeting recorder for iPhone. The model won't meet you halfway. The setup will.

Frequently asked questions