How accurate is AI transcription by language? Per-language WER benchmarks (2026)
AI transcription accuracy varies enormously by language. We compiled per-language word error rates from the Whisper paper FLEURS benchmark, showing which languages are near-human, which need editing, and which general models effectively cannot transcribe.
Transcription accuracy varies dramatically by language. On the FLEURS benchmark, OpenAI's Whisper achieves word error rates around 3-5% in English, Spanish, Italian, and German, in the high single digits for mid-resource languages like Dutch, Turkish, and Vietnamese, and above 75% in many low-resource languages such as Amharic, Yoruba, and Burmese, where a general model effectively cannot produce usable text. The gap comes down to training-data volume, phonetic complexity, and the writing system.
This guide presents per-language WER compiled from the Whisper paper's published FLEURS results (OpenAI, 2022), organized by accuracy tier. If you are evaluating transcription for a specific language, or wondering why your German audio transcribes flawlessly but your Thai audio does not, the data below explains the gap.
What these numbers are (and aren't). Every figure below is Whisper large-v2 WER on the FLEURS benchmark, from Appendix D.2.4 of the Whisper paper, the most complete per-language WER table any major lab has published. It is one model on one read-speech benchmark, not a live test we ran. Treat it as a comparable baseline, not current best-case: newer systems (Whisper large-v3, NVIDIA Canary, and language-specialized providers) score materially better on many of the harder languages. For example, large-v3 cuts Maltese from ~77% WER to roughly 26% character error rate. For Chinese, Japanese, Thai, Lao, and Burmese, the Whisper paper inserts spaces between characters and reports character error rate (CER), so those figures are already character-level, not word-level WER. Korean uses normal word spacing, so its figure is a true WER (inflated by Korean spacing conventions). Update, July 2026: we have since run our own reproducible benchmark of large-v3, large-v3-turbo, and small across 12 of these languages, with raw data published; where the two overlap, the tiers below hold up.
TL;DR: accuracy tiers at a glance
| Tier | WER range | Languages (representative) | What to expect |
|---|---|---|---|
| Tier 1 | under ~6% WER | English, Spanish, Italian, German, Japanese (CER), Polish, Russian | Near-human accuracy on clean audio |
| Tier 2 | ~6-11% WER | Dutch, Indonesian, Catalan, French, Turkish, Swedish, Malay, Vietnamese | Production-grade, minor edits needed |
| Tier 3 | ~11-16% WER | Thai, Greek, Czech, Croatian, Danish, Korean, Romanian, Mandarin (CER), Arabic | Usable, expect meaningful manual cleanup |
| Tier 4 | ~16-40% WER | Tamil, Hindi, Urdu, Hebrew, Serbian, Kannada, Marathi, Swahili | Rough draft quality, human review required |
| Tier 5 | ~45% to >100% WER | Nepali, Maltese, Yoruba, Khmer, Lao, Bengali, Burmese, Amharic | General models like Whisper effectively fail; specialized models required |
Numbers are Whisper large-v2 FLEURS WER (OpenAI, 2022). See the per-tier tables below for exact figures and the caveats above.
How WER benchmarks are measured
Understanding what each benchmark tests prevents the common mistake of comparing a lab score to real-world performance.
LibriSpeech (English only) uses clean audiobook recordings. It is the easiest common benchmark, so its numbers are the floor of what a model can do under ideal conditions. State-of-the-art English WER on LibriSpeech test-clean is around 1.4-2.7%.
FLEURS (Few-shot Learning Evaluation of Universal Representations of Speech) covers 102 languages with roughly 12 hours of speech per language, using the same sentences across languages (translations of Wikipedia/FLoRes content), which makes cross-language comparison meaningful. It is the most widely cited multilingual benchmark, and the source of the per-language numbers in this guide.
Common Voice (Mozilla) contains crowdsourced recordings across 100+ languages. Because speakers are non-professionals in varied environments, Common Voice WER is typically higher than FLEURS for the same language and model.
Real-world audio, with accents, overlapping speakers, background noise, and imperfect equipment, adds roughly 5-15 WER points on top of these benchmark numbers. A model reporting 5% on FLEURS may deliver 10-15% on a typical Zoom recording.
We re-ran the numbers ourselves in July 2026
To check whether the paper's large-v2 table still describes reality, we benchmarked three current Whisper checkpoints (large-v3, large-v3-turbo, small) on the same FLEURS test set: 50 utterances per language, 12 languages, 1,749 transcriptions, seed-fixed sampling, with the script and raw results downloadable. A few figures from large-v3 against the paper's large-v2 numbers below:
| Language | Paper large-v2 (below) | Our large-v3 (July 2026) | Reading |
|---|---|---|---|
| Spanish | 3.0% (Tier 1) | 2.9% WER | Agrees |
| English | 4.2% (Tier 1) | 4.1% WER | Agrees |
| Japanese | 5.3% CER (Tier 1) | 6.6% CER | Right at the Tier 1/2 line, within sample noise |
| Mandarin | 14.7% CER (Tier 3) | 6.7% CER | Halved; large-v3's biggest gain in our set |
| Hindi | 21.5% (Tier 4) | 15.7% WER | Improved into Tier 3 territory |
| Cantonese | not in large-v2 | 10.5% CER (turbo: 43.3%; small: no support) | New in the v3 generation, see note |
The tier structure held up: the strong languages stayed strong, and where our numbers moved, they moved in the direction the note above predicts (newer checkpoints scoring better on the harder languages, with Mandarin and Hindi the clearest examples). The Cantonese row is the one finding the paper's table cannot show you: support arrived with the large-v3 generation, so the same language swings from usable (large-v3) to broken (turbo) to unavailable (small) depending on the checkpoint. Full results, methodology, failure analysis, and downloads are in the 2026 benchmark article.
Tier 1: very high accuracy (under ~6% WER)
These languages have the largest training corpora and the most developer attention. Expect production-ready transcripts on clean audio with minimal editing.
| Language | Whisper large-v2 FLEURS WER | Notes |
|---|---|---|
| Spanish | 3.0% | Among the strongest in any benchmark |
| Italian | 4.0% | One of the best-covered European languages |
| English | 4.2% | Reference language; most benchmarks focus here |
| German | 4.5% | Strong on standard German; Swiss/Austrian dialects degrade |
| Japanese | 5.3% (CER) | Whisper measures Japanese at the character level; sentence quality is excellent |
| Polish | 5.4% | Exceptionally strong for a Slavic language |
| Russian | 5.6% | Good on standard Russian; regional accents degrade |
If you are transcribing in any of these on clean audio, the model matters less than the audio quality you feed it.
Tier 2: high accuracy (~6-11% WER)
Meaningful training data, but either less volume than Tier 1 or more phonetic complexity. Most production use cases work well; expect to correct occasional proper nouns and technical terms.
| Language | Whisper large-v2 FLEURS WER | Notes |
|---|---|---|
| Dutch | 6.7% | Benefits from proximity to German and English data |
| Indonesian | 7.1% | Strong for its resource level |
| Catalan | 7.3% | Punches above its speaker count thanks to dedicated datasets |
| French | 8.3% | European French dominates the training data |
| Turkish | 8.4% | Agglutinative morphology adds complexity |
| Swedish | 8.5% | Nordic corpora are well-curated |
| Ukrainian | 8.6% | Improved substantially post-2022 as datasets grew |
| Malay | 8.7% | Shares features with Indonesian |
| Vietnamese | 10.3% | Tonal; tone errors are the main failure mode |
Tier 3: medium accuracy (~11-16% WER)
This is where AI transcription becomes visibly imperfect. Transcripts are usable as a first draft, but expect to fix several errors per minute, especially around named entities and numbers.
| Language | Whisper large-v2 FLEURS WER | Notes |
|---|---|---|
| Thai | 11.5% (CER) | Character-level in the Whisper paper (no spaces between words) |
| Greek | 12.5% | Smaller training corpus than other European languages |
| Czech | 13.3% | Solid despite heavy morphology |
| Croatian | 13.4% | Shared South-Slavic features help |
| Danish | 13.8% | Difficult phonetics, but well represented |
| Romanian | 14.4% | Improving as datasets grow |
| Bulgarian | 14.6% | Moderate-resource Slavic language |
| Korean | 14.3% | Word-spacing inflates Korean WER; character-level accuracy is materially better, but the paper publishes no Korean CER |
| Mandarin | 14.7% (CER) | Character-level; higher than its reputation on FLEURS read speech |
| Arabic | 16.0% | Modern Standard Arabic; dialects (Egyptian, Levantine, Gulf) are much harder |
Mandarin's number surprises people: on clean, in-domain audio many tools do better, but on the standardized FLEURS character-level measure Whisper large-v2 lands in the mid-teens. Newer models narrow this.
Tier 4: low accuracy (~16-40% WER)
These languages often have hundreds of millions of speakers but limited labeled training data. Output is a rough draft that requires substantial human review.
| Language | Whisper large-v2 FLEURS WER | Notes |
|---|---|---|
| Tamil | 17.5% | Dravidian language with complex morphology |
| Hindi | 21.5% | High variance across accents and English code-switching |
| Urdu | 22.6% | Related to Hindi but in Perso-Arabic script |
| Hebrew | 27.1% | Right-to-left script, rich morphology |
| Serbian | 33.9% | Markedly harder than neighboring Slavic languages on FLEURS |
| Kannada | 37.0% | Dravidian family, underrepresented |
| Marathi | 38.3% | Indo-Aryan, moderate resources |
| Swahili | 39.3% | East-African lingua franca; dataset is growing |
For Tier 4, hybrid workflows (AI produces the first draft, a native-speaker editor cleans it up) are typically the highest-throughput option.
Tier 5: where general models effectively fail (~45% to >100% WER)
This is the tier the friendly-looking "~40-50%" numbers you see elsewhere get badly wrong. On FLEURS, Whisper large-v2 does not produce a rough-but-usable draft for these languages. It produces WER at or above 75%, often over 100%, which means more errors than words. They are effectively untranscribable by a general-purpose model without specialization.
| Language | Whisper large-v2 FLEURS WER | Notes |
|---|---|---|
| Nepali | 47.1% | Borderline; heavy review needed |
| Maltese | 76.6% | large-v3 improves this dramatically (~26% CER) |
| Yoruba | 94.8% | Tonal; large-v3 reaches ~49% CER, still hard |
| Khmer | 99.7% | Whisper large-v2 produces little usable output |
| Lao | 101.5% | Effectively untranscribable on large-v2 |
| Bengali | 104.1% | Huge speaker base, but Whisper large-v2 fails; large-v3 CER ~34% |
| Burmese | 115.7% | Effectively untranscribable on large-v2 |
| Amharic | 140.3% | More errors than words; needs a specialized model |
Two things matter here. First, newer and specialized models close much of this gap: Whisper large-v3, and providers who fine-tune on language-specific data, can turn an unusable large-v2 result into an editable one (Maltese and Bengali are good examples). Second, Sinhala is not in Whisper's FLEURS evaluation at all, so any specific "Sinhala WER" figure you see for Whisper is unsupported. Treat that language as untested by this benchmark.
What drives the accuracy gap
Training-data volume is the single strongest predictor. Whisper was trained on 680,000 hours of audio, the majority of it English; higher-resource languages get tens of thousands of hours, the lowest-resource ones a few hundred. More data sharply reduces WER, with diminishing returns as a language becomes well-covered.
Phonetic and writing-system complexity creates ceiling effects even with data. Tonal languages (Mandarin, Vietnamese, Thai, Yoruba) force the model to distinguish similar words by pitch. Agglutinative languages (Turkish, Finnish, Swahili) build long words from many morphemes, which interacts badly with tokenization. Logographic and space-free writing systems (Chinese, Japanese, Thai) shift the fair metric from WER to character error rate.
Audio-domain match matters as much as language. A model trained mostly on read-aloud audio underperforms on spontaneous conversation in the same language, which is why FLEURS read-speech numbers are a floor, not a promise for your meeting recordings.
How to improve accuracy for harder languages
Improve the audio first. Noise reduction, speaker isolation, and consistent levels can cut several WER points on real-world audio. Our noisy-audio guide covers the fastest wins.
Provide domain context. Many transcription APIs accept a list of proper nouns, technical terms, or phrases likely to appear; biased vocabularies cut substitution errors for jargon and named entities.
Choose the right model per language. Whisper leads on some languages, NVIDIA Canary or language-specific providers on others (notably Japanese, Korean, and Arabic). If a language is critical to your workflow, testing 2-3 providers on a representative sample is worth the hour.
Use a human editor for the last mile. For Tier 3 and below, a native-speaker editor reviewing an AI draft is several times faster than transcribing from scratch and lands final accuracy well above 98%.
Tools like Vocova transcribe across 100+ languages with automatic language detection, so you do not have to tag files by language in advance. Still, as the data above shows, the realistic accuracy you get depends on which language you are working in.
Frequently asked questions
Which language has the most accurate AI transcription?
English, Spanish, Italian, and German lead on the FLEURS benchmark, at roughly 3-5% Whisper WER on clean read speech, with Japanese close behind on character error rate. On real-world spontaneous audio, add 5-15 points to any of these.
How accurate is Whisper across languages?
On FLEURS, Whisper large-v2 lands under ~10% WER for roughly the Tier 1-2 languages in this guide, in the teens-to-thirties for Tier 3-4, and above 75% for a band of low-resource languages (Amharic, Yoruba, Khmer, Lao, Burmese, Bengali) where it effectively cannot transcribe. OpenAI has not published an equivalent per-language table for large-v3, so we measured one ourselves: our July 2026 benchmark covers 12 languages on large-v3, large-v3-turbo, and small, with raw data included.
What WER is considered "good"?
For most business uses, WER below 10% produces a transcript faster to read and edit than the audio; below 5% is near-human. Above 20% requires significant manual correction before it is usable as published text. See word error rate explained for how the metric is calculated.
Why is my German transcription more accurate than my Thai transcription?
German is a high-resource language (~4.5% FLEURS WER) with abundant data and features shared with English. Thai is tonal and written without spaces between words, with less labeled data, so its WER is roughly 11-12%, and the fair metric is character error rate. Even the best models keep a meaningful gap between the two.
Are FLEURS benchmarks comparable to real-world audio?
Not directly. FLEURS is clean, read rather than spontaneous, and professionally recorded. Real-world audio (meetings, phone calls, interviews) typically produces 5-15 points higher WER than FLEURS for the same language and model, so use these numbers to compare languages, not to predict your exact result.
Summary
AI transcription accuracy in 2026 is a function of language, audio quality, and model-task fit. On the Whisper FLEURS baseline, Tier 1 languages reach near-human accuracy on clean audio; Tier 3 needs editing; and a band of low-resource languages is effectively untranscribable by general models without specialization, far harder than the friendly double-digit numbers often quoted. Newer and language-specific models narrow these gaps, especially at the hard end.
If you are choosing a transcription pipeline, the most useful thing you can do is test your specific language and audio domain on 2-3 representative samples. Benchmarks rank languages; they do not predict your exact result.
Sources and further reading
- OpenAI, "Robust Speech Recognition via Large-Scale Weak Supervision" (Whisper paper, 2022): per-language FLEURS WER, Appendix D.2.4 / Table 13
- Conneau et al., "FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech" (2022)
- FLEURS-SLU (2025): Whisper large-v3 per-language CER, used here for the low-resource caveats
- Mozilla Common Voice datasets
- Word error rate explained
- How to transcribe audio in multiple languages
