Whisper Supported Languages and Accuracy by Language

Whisper supports 98 or 99 languages, depending on which source you count, and its accuracy differs far more than that number suggests. In the FLEURS benchmark, the large-v2 model from 2022 had a word error rate of 3.0 percent in Spanish ¹ and 94.8 percent in Yoruba ². The gap follows the training audio: languages with more hours of training data had fewer errors. Set the language yourself when you know it, and treat translation as a separate step that only goes into English.

Whisper supports 98 or 99 languages, but accuracy follows the hours of training audio

The count depends on the source. OpenAI's speech-to-text guide says Whisper supports 98 languages and that accuracy varies by language ³. Meta's paper on its MMS speech project describes Whisper as supporting the transcription of 99 different languages ⁴. Neither source explains the one-language difference, so treat the count as about 100.

The count matters less than the training data behind it. Whisper was trained on 680,000 hours of audio, and the authors report that 117,000 hours covered 96 languages other than English ⁵. The hours are very uneven from one language to the next, and that unevenness shows up in the test results.

The 2022 paper tested the model on FLEURS, a benchmark of 102 languages ⁶. FLEURS gives each language about 12 hours of speech for testing ⁶, which is enough to compare languages with each other, but it is a small, read-aloud test rather than a sample of everyday speech.

Whisper's training data has no speech at all for 20 of those 102 languages ⁷. So a language can appear on the supported list and still have little or no training behind it. Check the language you need against test numbers, as in the next section, rather than against the count.

The best-recognized languages are the ones with the most training data

The paper's main finding is a trend. The authors estimate that a language's word error rate halves for every 16-fold increase in its training data ⁸. The estimate is a fitted line with outliers, but its direction is clear: more training audio, fewer errors. A word error rate is the share of words that were wrong, counting substitutions, deletions and insertions, so lower is better.

The table lists nine languages from the same FLEURS test for the large-v2 model, sorted from the lowest error rate to the highest.

The chart uses the same numbers as the table. It also shows the shape of the trend: the languages at the top have the most support in the test, and the bars climb steeply once a language moves from Spanish and English toward the languages with the least data. Three points deserve care.

First, the paper scores some languages by character error rate instead of word error rate. Its list is Chinese, Japanese, Thai, Lao and Burmese, and Korean is not on it ¹⁹. Scores for those languages are character-based, so do not compare them directly with the word-based scores above.

Second, a word error rate above 100 percent is possible. It means the model inserted more words than the reference contains, so the error count is larger than the number of words that were spoken.

Third, the paper names Telugu, Hebrew, Chinese and Korean as the largest outliers from the trend. These languages have unique scripts and are more distantly related to the Indo-European languages that make up most of the training data ²⁰. The authors suggest linguistic distance, a poor fit of their tokenizer, or data quality as possible causes, and they did not test these.

OpenAI's own README makes the same point. It says Whisper's performance varies widely depending on the language, and it publishes a breakdown by language for the large-v3 and large-v2 models on Common Voice 15 and FLEURS ²¹. The OpenAI Whisper repository has the chart.

These numbers come from read sentences in a benchmark, not from everyday dictation. Accents, background noise and specialist vocabulary raise error rates in every language. The guide to speech recognition accents covers that gap in more detail.

Whisper detects the language from the start of the audio, and can guess wrong

If you do not set a language, the open-source code decides it from the audio. The transcribe function detects the spoken language from up to the first 30 seconds of audio ²². Nothing in that step re-checks the language later, so a recording that switches language after the opening can be transcribed with the wrong language for the rest.

Detection is good but not perfect. On the FLEURS language identification test, the best Whisper model named the right language 80.3 percent of the time, across the 82 languages it had training data for ²³. The paper says the model was heavily penalized for the 20 languages with no training data, so the benchmark score reflects missing data as well as model quality.

Speech that mixes languages is harder still. On two conversational corpora of code-switching, where speakers move between two languages, Whisper's mixed error rate on the code-switched utterances was about 39 percent ²⁴. The same study found that recognition significantly altered and reduced the speakers' original code-switching.

The fix is to name the language. The Whisper README gives this command-line example for Japanese audio: whisper japanese.wav --language Japanese ( README on GitHub). If the output looks like a different language from the one you spoke, check the language setting first. The question on why Whisper transcribes in the wrong language covers the common causes.

Its translation mode works only into English

Whisper can translate speech, but the direction is into English. The paper's training data includes 125,000 hours of speech translated into English, described as X to English translation ²⁵. The README describes its translate task as translating speech into English, with the example --task translate for Japanese audio.

The turbo model is the exception to check. The repository says turbo is not trained for translation, so to translate non-English speech into English you should use one of the multilingual models instead ²⁶. The README adds that turbo returns the original language even when the translate task is set, so if you asked for English and got the original language back, switch to a multilingual model.

If you need a transcript in the original language and an English version, run the transcription and the translation as two steps. For interviews, lectures or calls that a team reads in English, keep the original-language transcript as the record and create the English version for readers who need it. That keeps the speaker's words intact and makes any translation error easy to trace back to the source. The question on whether Whisper can translate audio explains the two tasks, and the guide to translating audio to English covers the tools that do this on a phone or computer.

Fine-tuned versions improve single languages such as Japanese and Hindi

Fine-tuning means training a Whisper model further on one language's audio. The gains can be large, and they show up in the same measurements as the figures above.

For Japanese, a 2024 study found that the untuned Whisper-Tiny model had a character error rate of 32.7 percent, and that fine-tuning on Japanese data lowered it to 14.7 percent ²⁷. Tiny is the smallest Whisper model, so the absolute numbers are higher than the large models would show. The gain from Japanese training is the lesson. Character error rate counts individual characters rather than words, which suits Japanese, where transcripts may or may not include spaces between words ²⁸. The study measured its own Japanese test sets, so its figures describe that study rather than every Japanese recording.

For Hindi, the Vistaar study fine-tuned Whisper for 12 Indian languages on 10.7K hours of public audio, which it called IndicWhisper. Across the Hindi subset of the Vistaar benchmark, IndicWhisper averaged 13.6 percent word error rate across seven test sets ²⁹. On the FLEURS part of that benchmark, it scored 11.4 percent ³⁰, against 21.5 percent for the untuned large-v2 model on the same FLEURS test ¹².

Read the table with care. The Japanese rows are character error rates on a different test, and the Hindi FLEURS rows are the closest like-for-like pair, though the Vistaar authors' text normalization may differ from the Whisper paper's. A fine-tuned model is a separate model you choose, so check which one an app uses before you assume the Hindi result applies. The guide to Hindi voice typing and the question on how good Whisper is at Hindi give the wider picture. For Japanese, see the guide to Japanese, Chinese and Korean speech to text.

Newer multilingual models from other labs cover more languages

Several newer models cover far more languages than Whisper. Meta's MMS project built a single speech recognition model for 1,107 languages ³¹. Its paper compares it with Whisper on FLEURS-54, the 54 languages both models support, and reports that MMS reduced Whisper's word error rate by a relative 58 percent while supporting over 11 times as many languages ³². That comparison is the authors' own, and the two models used different decoding setups.

Meta's later Omnilingual ASR covers more than 1,600 languages ³³ and is licensed under Apache 2.0 ³⁴. Its paper reports that the 7B model meets a character error rate threshold of 10 percent for 78 percent of the 1,570 languages it was evaluated on ³⁵. The same paper shows the pattern this guide has followed throughout: only 36 percent of the languages with less than 10 hours of training audio meet that threshold ³⁶, which is why low-resource languages lag behind.

A higher count does not guarantee accuracy in your language. Each source reports its own test sets, so a language list from one vendor cannot be compared line by line with another. The low-resource figure is the one to keep in mind for less common languages, because it shows how many of them still fall short of the threshold even in the newest model family.

When you compare models, check four things: which languages the test covers, whether the test is read speech or conversation, whether the score is a word or a character error rate, and whether the model is the one your app actually runs. A score that answers all four questions is one you can use. The guide to dictation languages across apps and devices compares which languages the tools you use actually offer.

What this means when you pick a language setting

• Set the language when you know it. Auto-detection looks at the first 30 seconds only, and it can be wrong.
• Expect the error rate to follow the data. A language with more training audio, such as English or Spanish, is transcribed more accurately than one with little, as the trend in the paper shows.
• Test with your own speech. Dictate a few minutes in your accent and vocabulary before you rely on a language.
• Use translation as a second step into English, with a model that supports translation.

For dictation on a Mac or Windows PC, Yap, which we make, runs on the device. For the languages it supports, check the list on our download page rather than this guide, since the lists change between releases.