Is speech recognition less accurate in languages other than English?

Often, yes. Some languages are recognized far less accurately than English, and in some the gap is very large. English is not the reference point either. In one benchmark, English was only slightly worse than Spanish, while Yoruba was far behind. The size of the gap differs from one language to the next.

Accuracy varies widely from one language to the next

OpenAI says that Whisper's performance "varies widely depending on the language" ¹. The clearest numbers come from the FLEURS benchmark, a shared test set that lets the same speech be scored in many languages. On that test, Whisper large-v2's word error rate in Yoruba was about 31.6 times its rate in Spanish ². In English it was about 1.4 times its Spanish rate ³. So its Yoruba error rate was far higher than its Spanish or English rates.

A single benchmark does not cover every language or every tool. It does show that the gap is not simply English against everything else.

Data coverage is very uneven across languages

A 2025 review of code-switching speech recognition found that a small subset of languages receive most of the attention, and that the languages covered by datasets are very uneven ⁴. A 2025 study of isiXhosa and Afrikaans preschool narratives says Whisper's training data includes no isiXhosa and no closely related languages ⁵. Its authors chose Swahili as the language token when they fine-tuned on isiXhosa.

For the search-level view of which languages have support at all, see the guide to Whisper's supported languages. The glossary entry on low-resource languages defines the term.

Accent and mixed-language speech add a second gap

Accented speech can add errors on top of the language gap. A 2007 study of foreign-accented English found that the best combination of pronunciation modeling and acoustic adaptation cut word error rate by 46 to 71 percent relative ⁶. Speech that mixes two languages adds another problem. The authors of a 2023 paper say that multilingual models such as Whisper often struggle with code-switching ⁷. For more on accents, see the guide to speech recognition accuracy by accent.

The vendor figure covers only the top languages it tested

OpenAI says word error rate in its ChatGPT dictation evaluations was at least 10 percent lower for the top languages it tested than with the previous production model ⁸. That is a claim about the top languages in its test. It says little about a smaller language, so do not read it as a promise for yours.

It depends on the language, the tool and your speech

The same language can perform very differently across tools. Results also depend on your microphone, background noise and the vocabulary you use, so a benchmark is a guide, not a guarantee. If a language matters to you, test the tool on your own speech, in the conditions where you will use it.

What to do

1. Check the vendor's own language list for your exact language and region, then check whether it works offline. The guide to dictation languages sets out those checks.
2. Dictate five minutes of your own writing in the language, then count the errors. The glossary entry on word error rate explains how to count them.
3. If the error rate is high, try a second tool on the same sample, because the engine can matter as much as the language.
4. If you switch between languages, test that on purpose, since mixed speech is where many tools fail.
5. Pick the tool that makes the fewest errors on your sample, not the one with the longest list.

For related questions, see which dictation app supports the most languages, how many languages Google Speech-to-Text supports and words per minute by language.