What is WER (word error rate)?

WER counts the minimum number of word substitutions, deletions, and insertions needed to turn the transcript into the reference text, divided by the number of words in the reference. A WER of 3% means roughly 3 mistakes per 100 words. Lower is better. For languages without spaces between words, such as Chinese and Japanese, the same calculation is done per character instead.

Why bands instead of exact numbers?

Each band is a truthful coarsening of a published result: the published figure for every language shown falls inside its displayed band. The bands are not new measurements or estimates. We publish bands rather than exact third-party figures, and we plan to publish exact numbers once we have run our own evaluation through the production pipeline.

Did transcribe.so measure these results?

No. The bands come from published benchmark results, not from our own measurements. We never estimate a result that has not been published, and we show "Not published" instead of a guess.

Which models does transcribe.so actually run?

Every file runs on our Standard engine, the production speech-to-text stack whose published results this page summarizes. Speaker labels and word-level timestamps are included on every transcript, with no model to pick.

Accuracy benchmarks

Transcription accuracy by language

Published error-rate bands for the engine transcribe.so ships, per language. Each band is a truthful coarsening of a published benchmark result: the published figure for every language shown falls inside its displayed band. Nothing on this page is estimated, and we say "Not published" rather than guess.

Last updated 2026-07-28. Jump to: Korean · Japanese · Chinese (Mandarin) · English · Methodology · Limitations

How to read this page

WER (word error rate) is the standard accuracy metric for speech-to-text. It counts wrong, missing, and extra words against a human-verified reference transcript, so lower is better. For languages written without spaces, such as Chinese, Japanese, and Cantonese, the same count is done per character in the cited evaluation. We group each published result into one of three bands: Under 5%, 5% to under 15%, and 15% or higher. The same boundaries apply to character error rate for the languages scored per character. Bands are comparable within a language, not across languages.

Error-rate band by language

FLEURS is a public multilingual benchmark of read speech built by Google Research. The table lists every language our Standard engine, the default engine on transcribe.so, supports, in alphabetical order, with the band its published FLEURS result falls into. The bands summarize published results; they are not our own measurements.

Language	Published error-rate bandFLEURS, lower is better
Arabicالعربية	5% to under 15%
Cantonese粵語	Under 5%
Chinese (Mandarin)中文	Under 5%
CzechČeština	15% or higher
DanishDansk	5% to under 15%
DutchNederlands	Under 5%
EnglishEnglish	Under 5%
FilipinoFilipino	15% or higher
FinnishSuomi	5% to under 15%
FrenchFrançais	Under 5%
GermanDeutsch	Under 5%
GreekΕλληνικά	5% to under 15%
Hindiहिन्दी	5% to under 15%
HungarianMagyar	15% or higher
IndonesianBahasa Indonesia	Under 5%
ItalianItaliano	Under 5%
Japanese日本語	Under 5%
Korean한국어	Under 5%
MacedonianМакедонски	Not published
MalayBahasa Melayu	5% to under 15%
Persianفارسی	15% or higher
PolishPolski	5% to under 15%
PortuguesePortuguês	Under 5%
RomanianRomână	5% to under 15%
RussianРусский	Under 5%
SpanishEspañol	Under 5%
SwedishSvenska	15% or higher
Thaiภาษาไทย	5% to under 15%
TurkishTürkçe	5% to under 15%
VietnameseTiếng Việt	Under 5%

Bands coarsen results from the engine's published technical report on the FLEURS benchmark (Google Research). Languages are sorted alphabetically, never by result, and we do not publish ordering within a band. Where a same-split comparison has been published, the engine we ship shows 1.5x-2.3x fewer errors than Whisper large-v3 in published same-split comparisons.

See it in action

Bands are one thing. See the actual output.

Error rates tell you how many words land. This is what a finished transcript looks like: chapters, speaker-aware sections, and searchable Q&A on a real recording.

The full-length interview with Elon Musk | The Economist

The Economist

Try this real transcript

Who said what, when

Contents✍️Human written like chapters

20 chapters · 60 sections

1The Age of Abundance is Coming

2Why You Won't Control Superintelligence

3AI Doom is Inevitable (Eventually)

4How Rivals Keep AI Honest

5China's AI Rise Changes Everything

Ask this video

🎯Get who said what, when

Answer

Sooner than most people expect: Musk guesses AI may exceed the sum of human intelligence in roughly five years, and be far greater than it within ten . He says it is unlikely humans will still be in control ten years from now, comparing the gap to the one between humans and chimpanzees . And he admits progress looks so inexorable that he could not stop it even if he wanted to .

Language details

Korean한국어

Korean sits in the strongest published band, under 5% error rate on FLEURS. That is the territory where a transcript of clean speech is usable without a correction pass.

OpenAI has not published per-language FLEURS results for Whisper large-v3, so we show no Whisper band for Korean rather than estimating one. Korean services built on Whisper-family models inherit whatever Whisper does on Korean, and none of the Korean transcription products we are aware of publish a measured error rate at all.

On transcribe.so, Korean audio runs through the Standard engine by default. Word-level timestamps are supported, and diarization is included in the base price.

Japanese日本語

Japanese also falls in the under 5% band on FLEURS with the Standard engine, which every Japanese file on transcribe.so runs through by default.

Japanese error rates are computed at character level because Japanese text has no whitespace word boundaries. Character-level and word-level percentages are not directly comparable across languages; compare models within a language instead.

As with Korean, OpenAI publishes no per-language FLEURS figure for Whisper large-v3 on Japanese, so no Whisper comparison is shown. We do not estimate results a vendor has not published.

Chinese (Mandarin)中文

Chinese is one of the few languages with a clean same-split Whisper comparison. Across the languages where such comparisons exist, the engine we ship shows 1.5x-2.3x fewer errors than Whisper large-v3 in published same-split comparisons. Mandarin sits in the under 5% band.

Cantonese is also in the under 5% band for our Standard engine, and it is one of the languages covered by the same published same-split comparison. Whisper-derived tools inherit a substantially higher published error rate there.

Chinese dialect and accent coverage is included as part of the 52 languages and dialects transcribe.so supports.

EnglishEnglish

English has a clean same-split comparison: on FLEURS read speech, the engine we ship shows 1.5x-2.3x fewer errors than Whisper large-v3 in published same-split comparisons, and English sits in the under 5% band.

FLEURS is read speech recorded in quiet conditions. Expect absolute error rates on real-world recordings such as meetings, earnings calls, and noisy web audio to be higher for every model; the relative ordering between models is the informative part.

Methodology

Where the bands come from. We did not run these benchmarks ourselves. Each band is a truthful coarsening of a published result: we take the published per-language figure and report only which of the three ranges it falls into. The band is not a new measurement or estimate, and the published result for every language shown falls inside its displayed band. When no result has been published for a language, the row says so.

FLEURS. A public benchmark from Google Research covering 102 languages. Each language's test set is read speech: sentences from the FLoRes translation corpus recorded by native speakers, a few hours of audio per language. Read speech is recorded under cleaner conditions than meetings, lectures, podcasts, and other real-world audio, so treat the bands as a best case.

Scoring. Figures behind the bands are word error rates, except for languages without whitespace word boundaries (Chinese, Japanese, Cantonese), which are scored per character in the cited evaluation, as is standard practice. WER also depends heavily on how text is normalized before scoring: lowercasing, punctuation removal, number and abbreviation expansion. Normalization differences alone can move WER by whole percentage points, which is one reason accuracy claims without stated normalization rules are not comparable, and why we do not compare results across languages.

Comparisons. We only compare models when the figures come from the same published evaluation on the same splits. Where such comparisons exist, the engine we ship shows 1.5x-2.3x fewer errors than Whisper large-v3 in published same-split comparisons. We do not publish per-language ratios or estimate results a vendor has not published.

What transcribe.so ships. Every file runs on our Standard engine, the production stack summarized above, with speaker labels and word-level timestamps on every transcript. Diarization is included in the base price, not an add-on.

Honest limitations

FLEURS is read speech recorded in quiet conditions. Real meetings, lectures, and podcasts are noisier and more spontaneous, so absolute error rates on your audio will be higher for every model, and a band on this page does not predict the error rate on your recording.
The table leans on one corpus. A model can land differently on other test sets, accents, or domains.
Bands deliberately discard precision. Two languages in the same band can have meaningfully different published results; we accept that loss rather than publish exact third-party figures as if they were our own.
Whisper large-v3 has no vendor-published per-language FLEURS results for most languages, including Korean and Japanese. We show no comparison there instead of estimating one.
Benchmarks do not cover heavily accented speech, domain jargon (medical, legal), overlapping speakers, or very low-quality recordings. No public benchmark we are aware of does this well across languages.
Accuracy claims you may see elsewhere, such as "99% accurate", are usually published without a test set, normalization rules, or model version, and cannot be compared to anything on this page.

Test it on your own audio

The only benchmark that matters is your recording. Free credits included, no credit card required, and you see the exact price before you confirm.

Start free See pricing

Works on YouTube links, podcasts, meetings, lectures, and voice memos.