How do you translate a voice recording into English text?

The dependable route is two steps. First, transcribe the recording in the language it was spoken in. Second, check that transcript against the audio and translate the checked text into English. Some tools hide both steps behind one button. That is convenient, but a wrong word in the transcript becomes a wrong word in the English, and fluent English can hide it. For the basics, see our overview of transcription, and for the detail on other languages, see our guide to transcribing and translating audio in other languages.

Some tools output English straight from the audio, but only some models can do it

OpenAI's speech-to-text guide names whisper-1 as the model for translating a completed audio recording into English ¹. If you run Whisper yourself, the command-line tool has a translate option, and the README's example turns Japanese speech into English with the medium model ².

The turbo model does not translate. The README says it is not trained for translation tasks, so translating non-English speech into English needs one of the multilingual models instead ³. Check which model a tool runs before you trust its English output.

Buzz describes itself as transcribing and translating audio offline on a personal computer ⁴. Google's Gemini audio documentation lists transcription and translation of speech among the things it does with audio input ⁵. Because that output is generated text rather than a fixed transcript, the check below still applies.

Whisper's translate task produces English. A different target language needs a second, separate translation step on the transcript.

Language, audio quality and language detection decide whether the English can be trusted

Accuracy depends on the language. In a published benchmark, Whisper large-v2 had a word error rate of 94.8 percent in Yoruba ⁶. At that level almost every word in the transcript is wrong, and a translation built on it would carry those errors into English. Read sentences in a controlled test are the easiest case, so real recordings are usually harder. If your recording is Hindi or Hinglish, read the question on transcribing Hindi audio before you start.

Language detection is the other weak point. If you leave the language unset, the tool has to guess it. In openai-whisper the language is unset by default, and the program then detects the spoken language itself ⁷. A wrong guess can affect the whole transcript, so set the language yourself whenever you know it.

Recordings that switch between two languages inside a sentence are the hardest case. A 2025 study says spontaneous switching causes language confusion for code-switching recognizers ⁸. Choose the language most of the speech is in, then read the points where the language changes, since errors cluster there. For the term, see code-switching.

Check the English against the recording, in this order

1. Set the spoken language, then transcribe the whole recording.
2. Read the first minute or two of the transcript against the audio. If the words are in the wrong language or script, change the setting and run it again.
3. Correct names, numbers and dates against the audio. A wrong number or name changes the meaning of a sentence, so fix these first.
4. Translate the corrected transcript, not the raw one.
5. Read the English against the transcript, sentence by sentence, looking for a change in meaning rather than word-for-word matches.
6. Have a fluent speaker check anything that matters, such as a contract, a medical instruction or a legal statement.

Word error rate, the usual score for transcripts, does not work well for translations. It compares a transcript with one correct reference, and a sentence can be translated correctly in several wordings. So step 5 is a reading check, not a score. A practical version is to ask a second person who speaks the language to translate a few key sentences back into the spoken language and compare them with the transcript. A large change in meaning needs fixing before you use the text. For the score itself, see word error rate.

When a one-step tool is enough, and when it is not

For a casual voice note in a language you can read, a one-step tool is often enough, as long as you can check the English yourself. For a recording that will be quoted, filed or acted on, do the two-step check. Use a professional human translator for legal or medical content, because a small error in those texts can matter a great deal.

If you only need the transcript first, our free audio to text tool transcribes an uploaded file in your browser. For the recording side of a voice note, see voice memos and voice notes to text.

For related questions, see how to transcribe Japanese, Chinese or Korean audio and whether AI can transcribe African languages.