How Accurate Is AI Transcription, By Language?

No speech-recognition system available today reaches 99% word accuracy on a real field recording, in any language. That is not a gap in one particular tool — it is the current state of the technology. The honest promise a transcription product can make is different: 99%+ accuracy in the transcript you export, reached by an AI draft plus your review, not by the machine alone.

Why published accuracy numbers don't match your recording

Vendors publish word error rates measured on clean, single-speaker, scripted audio recorded on good equipment. That is not what a field interview sounds like.

Real interviews have crosstalk, background noise, phone microphones, regional accents, dialect, and participants switching languages mid-sentence. Every one of those raises the error rate, and they compound on top of each other. A benchmark number and a 30%+ real-world error rate on the same underlying technology can both be true — they just describe different audio.

What "accurate transcription" should actually mean

Treat a machine transcript as a fast first draft, not a finished research record. The output is only as trustworthy as the review it gets. A transcript workspace that keeps the audio, timestamps, and text side by side — so a reviewer can jump to the exact moment behind any uncertain line — is what turns a draft into something you can quote.

See the review workspace

Does accuracy vary by language?

Yes, and by a wide margin. As a rough guide:

• Well-resourced languages (widely spoken languages with large amounts of training data — English, Spanish, French, and similar) tend to see roughly 85-95% word accuracy on decent field audio. A light review pass is usually enough.
• Other supported languages — many South Asian, Southeast Asian, African, and Eastern European languages among them — tend to land around 60-80%. Still far faster than typing from scratch, but the output is a draft, not a finished result, and deserves a fuller review pass.

Two things worth knowing about these ranges: they're expectations, not a guarantee for your specific recording — actual accuracy depends heavily on your audio, not just the language — and for most languages they're extrapolated from published clean-speech benchmarks rather than measured directly on field recordings. We don't publish an exact percentage for your language, because the number that matters is the one your own recording gets, and any single figure we print would be wrong for someone.

Why dialect — not the language — is usually the real ceiling

The accuracy range above describes a language's standard form. What a participant actually speaks is often not that, and the gap between the two is where most remaining errors live.

The typical failure mode is a model quietly normalizing colloquial speech toward the more common, standard-sounding version of a word or phrase — a dialectal number replaced with an unrelated common word, an address term collapsed into a similar-sounding one, a colloquial phrase dropped entirely. This is the dangerous kind of error, because the output still reads fluently. Nothing about it looks broken, so it's easy to miss without checking against the audio.

Where a study depends heavily on dialectal or colloquial speech, budget time for a full review pass rather than expecting the first draft to carry it.

What we deliberately don't do to inflate the numbers

A few things that looked like easy wins on paper and didn't hold up when tested properly, so we don't ship them:

• Audio enhancement or denoising. Cleaning up noisy audio before transcription sounds like it should help. Tested, it didn't — modern speech models are already fairly noise-robust, and stripping acoustic detail to "clean" the signal can remove information the model was actually using. This matches independent published research on the same question.
• Keyword-boosting. Supplying a list of expected terms can look like it improves consistency, but it does so by making the model more confident, not more correct — including confidently wrong. A visible gap in a transcript is safer for research than a confident invented word.
• Chunk-length tuning. Splitting audio into shorter or longer windows before transcribing tested out as roughly neutral, not a real lever.

None of these change what we recommend to users: better source audio and a real review pass, not a processing trick.

Before you record: a few things that help

There's no hardware requirement here — record on the phone you already have. But a few free habits measurably help the draft you get back, without asking anyone to buy new equipment:

• Record somewhere quieter when that's practical for the interview.
• Keep the phone or recorder close to whoever is speaking, rather than across the room.
• Avoid a pocket or a surface that vibrates — a table someone might tap or lean on.
• Do a short test clip and listen back before the real session starts.
• If the setting allows it, keep overlapping speakers from talking at the exact same time.

None of this is required, and none of it is ever used to block a job — audio-quality warnings are advisory, not a gate. It only ever helps at the margin. Reviewing the transcript against the audio still matters regardless of how the recording sounds.

How to get the most accurate transcript for your study

1. Set your project's expected languages. This tells the product what to expect instead of guessing, and for code-switching interviews, list every language participants actually use.
2. Fill in the glossary. Local names, place names, and domain terms are the single highest-leverage thing you can give the system before uploading.
3. Use the review workspace as the main event, not a fallback. Segment-level playback and click-to-seek are how a transcript goes from a rough draft to something you can put in a report.

Start transcribing

See Bengali interview transcription for a language-specific walkthrough, or the broader research interview transcription guide for the full workflow.