Transcribe and Translate Audio in Other Languages

AI transcribers handle widely spoken languages well and many others poorly, and the gap can be measured. In a published benchmark, OpenAI's Whisper large-v2 had a word error rate of 3.0 percent in Spanish ¹ and 94.8 percent in Yoruba ². Translation into English is a separate step on top of transcription, and speech that mixes two languages is especially hard. To transcribe a recording in another language, set that language yourself, check the transcript in the language you chose, and check any translation as its own text. This guide shows the published numbers, what they do and do not cover, and the checks that catch the common failures. For one language at a time, start with the question on Hindi or the question on African languages.

Transcription accuracy depends heavily on the language, and published benchmarks show by how much

Accuracy is measured as word error rate, the share of words in a transcript that are substituted, deleted or inserted compared with a correct reference. A widely cited comparison by language comes from the paper that introduced Whisper. Its authors tested Whisper large-v2 on FLEURS, a set of read sentences recorded in many languages, and reported the error rate for many languages in a table.

The results fall into clear bands. English and Spanish are near the bottom of the error scale, at 4.2 percent in English ³ and 3.0 percent in Spanish ¹. Arabic is at 16.0 percent ⁴ and Hindi at 21.5 percent ⁵. Swahili reaches 39.3 percent ⁶. Yoruba is at 94.8 percent ² and Telugu at 99.0 percent ⁷, which means that almost every word in those transcripts counted as an error.

The gap between the best and the worst of these languages is very large. Yoruba's error rate was about 31.6 times Spanish's ⁸. The Whisper authors estimate that a language's word error rate halves for every 16-fold increase in its training data ⁹. Their training set held 680,000 hours of audio, and 117,000 hours of it covered 96 languages other than English ¹⁰. OpenAI says the model's performance varies widely depending on the language ¹¹, so check the results for your language before you choose a tool.

Newer systems change the ranking, and not in one direction

The 2022 table is not the latest picture. In 2025, Mistral AI's Voxtral paper compared Whisper large-v3 with GPT-4o mini Transcribe and Gemini 2.5 Flash on the same FLEURS test, language by language. The three systems do not rank the same way in every language.

Word error rate on the FLEURS read-speech test, from Table 4 of the Voxtral paper. Lower is better.

In Hindi, Gemini 2.5 Flash had the lowest error rate of the three. In Arabic, it had the highest. Whisper large-v3 was the weakest of the three in Hindi. Its predecessor, large-v2, scored 21.5 percent on the 2022 Hindi test ⁵, while large-v3 scored 28.87 percent in 2025 ¹². The two numbers come from different papers and different evaluation runs, so read the difference as a warning that a newer model is not automatically better in every language, not as a measured change between versions.

The choice of test set matters too. In the same Voxtral paper, Whisper large-v3 scored 46.75 percent on Hindi in Mozilla Common Voice ¹⁸, against 28.87 percent on FLEURS ¹². Common Voice is built from recordings that many volunteers contribute ¹⁹, so its audio varies more than a controlled benchmark does. Your own recordings may resemble Common Voice more than FLEURS, although that depends on where you record. Treat the FLEURS numbers as a best case.

Set the spoken language yourself rather than relying on automatic detection

Set the spoken language yourself whenever the tool lets you. When the language is left unset, the tool has to work it out. In openai-whisper, for example, the language is unset by default and the program then detects the spoken language itself ²⁰. Detection is convenient, but a tool that is told the language does not have to guess it, and a wrong guess can affect the whole transcript.

We did not find a published accuracy figure for language detection on short clips, so this guide does not put a number on that risk. The practical rule follows from the design. Tell the tool the language when you know it, and check the first part of the output when you do not.

On a Mac, Apple's Mac Help says you can change the language you dictate in, or set up Dictation for multiple languages and switch between them as needed ²¹.

1. Choose the spoken language before you start. If one recording mixes languages, choose the language that most of the speech is in.
2. Transcribe the first minute or two and read it. If the words look like the wrong language or the wrong script, change the language and run it again.
3. Check the rest of the transcript against the audio at several points, not only at the start. Errors often appear later, when the speaker changes topic or speed.
4. Only then translate, and check the translation as its own text, as described below.

If you dictate live on a computer rather than transcribing a file, Yap, which we make, is an on-device dictation app for Mac and Windows. For its current language list, see the download page. To transcribe a recording file in a browser, our free audio to text tool works from the file you upload.

Mixed-language speech such as Hinglish is a hard case for AI transcribers

Speech that switches between languages inside a sentence, which linguists call code-switching, is a hard case for AI transcribers. Hinglish, which mixes Hindi and English in one conversation, is a common example. The problem is partly data. A 2025 review found that a small subset of languages receive most of the attention, and that there is a large disparity in the languages the datasets cover ²².

A second problem is that recognizers can confuse the two languages. The authors of a 2025 study on code-switching recognition say that spontaneous switching within a sentence causes language confusion ²³. They also say that even when each language in a mix is well resourced, the shortage of annotated code-switched audio makes recognition harder ²⁴. The shortage is not new. A 2018 Hindi-English corpus paper said that very limited code-switched resources were available at the time ²⁵.

Apple's Dictation help describes switching between languages as needed, and it says nothing about a single sentence that mixes two languages ²¹. If you set Dictation to one language and speak a mixed sentence, check the output for the words from the other language first.

We did not find a peer-reviewed, public Whisper error rate for Hinglish. Published Hindi-English results use different test sets and metrics, so they cannot be compared with the FLEURS figures above. Treat the Hindi numbers in this guide as the best case for Hindi alone, and expect more errors wherever English words appear. In practice:

• Pick the dominant language of the recording and check the transcript for English words that came out in the wrong form.
• Read the switch points closely, because that is where errors cluster.
• For a language-specific walkthrough, see the Hindi question and the guide to Hindi voice typing.

Translating a recording is transcription first, then translation, even when one tool does both

A translation of a recording is two steps. First the speech becomes text in the spoken language. Then that text becomes text in another language. Some tools hide the two steps behind one button, but the errors from the first step carry into the second, so you should check both.

In OpenAI's API, the translation option produces English text. Its speech-to-text guide names the whisper-1 model for translating a completed audio recording into English ²⁶. The Whisper paper describes one model that performs several tasks on the same input audio, transcription and translation among them ²⁷. If you need a different target language, a transcript followed by a separate translation step is a common route.

A wrong word in the transcript becomes a wrong word in the translation, and fluent English can hide the error. That is why the checks should come in this order:

1. Read the transcript in the spoken language and correct names, numbers and dates against the audio.
2. Translate the corrected transcript, not the raw one.
3. Read the translation against the corrected transcript, sentence by sentence, looking for changed meaning rather than word-for-word matches.
4. Have a fluent speaker check any passage that matters, such as a contract, a medical instruction or a legal statement.

Translation is harder to score than transcription. Word error rate compares a transcript with one correct reference, but a translation can be correct in several different wordings, so a single error count does not work well for it. A practical check is to ask a second person who speaks the spoken language to translate a few key sentences back, then compare the result with the original transcript. A large change in meaning is a problem to fix before you use the text, even when the English reads smoothly.

For a step-by-step version of the English case, see the question on translating a voice recording to English and our guide to translating audio to English. For the general term, see speech translation.

Scripts without spaces and right-to-left scripts need language-aware checking

Word error rate assumes that you can tell where one word ends and the next begins. In some writing systems that is not obvious. Chinese and Japanese text has no spaces between words, and the boundaries are a matter of judgment. A study of Mandarin speakers found that their agreement on where the word boundaries fall was about 75 percent ²⁸, so a word count for Chinese depends on a boundary choice that even native speakers do not fully agree on. The character error rate avoids that choice by counting characters instead of words. If a tool reports a word error rate for Chinese or Japanese, ask what it counted as a word. For the specific guides, see the question on Japanese, Chinese and Korean audio and our guide to Japanese, Chinese and Korean speech to text.

Right-to-left scripts such as Arabic and Hebrew have a different problem: the text mixes directions. The Unicode standard says that when right-to-left scripts use digits written from left to right, the text is actually bidirectional, a mixture of right-to-left and left-to-right text ²⁹. A transcript in Arabic often also contains numbers, English names and Latin-script brand names, so check the order of those items after you paste the text into another app.

Accuracy in Arabic is also measurable. Whisper large-v2 had a word error rate of 16.0 percent in Arabic on the FLEURS test ⁴, and the newer comparison above shows that the ranking of tools changes for Arabic too. When you check an Arabic transcript:

• Read it in its own direction, then check the numbers and Latin-script names separately against the audio.
• Read the first and last words of each line, where direction problems are easiest to spot.
• Use the Arabic speech to text guide for the setup details, and see the right-to-left script term for the wider issue.

For low-resource languages, test on your own audio and plan for a human review

A low-resource language has little training audio, and a published accuracy figure for it may not apply to your recordings. Meta's Omnilingual ASR announcement is one example of the gap. Meta says its speech recognition system reaches character error rates below 10 for 78 percent of the languages it covers ³⁰. That is a share of languages under a threshold. It does not tell you how accurate the system is for any single language.

The Whisper authors' estimate implies that each 16-fold drop in a language's training data roughly doubles its error rate ⁹. Another failure does not show clearly in a word error rate. The Whisper authors list complete hallucination among the known failure modes, in which the model outputs a transcript entirely unrelated to the actual audio ³¹. A transcript can look fluent and still be wrong, so read the whole output and not only the parts that look right.

For a language like this, run your own test before you rely on a tool:

1. Choose three to five recordings that represent your real audio, including the hardest ones: fast speech, background noise and any mixing of languages.
2. Transcribe each one with the language set, as described above.
3. Correct one of them by hand, word by word, to make a reference transcript.
4. Count the words that are wrong against that reference, which gives you the word error rate for your own audio.
5. Decide what a wrong word would cost. If an error could change the meaning, plan a human review or a human transcriber for that work. Human and AI transcription compared sets out the trade-offs.

For African languages in particular, the question on whether AI can transcribe them covers which languages are covered at all. The guide to Whisper's supported languages lists what the model was trained on, and the term for a low-resource language explains the category.

For the wider picture of how transcripts are produced and reviewed, start with our overview of transcription, then how to transcribe audio to text, the Whisper, ChatGPT and Gemini options and how to pick transcription software. For a voice note in another language, the guide to voice memos and voice notes covers the recording side.