Transcribe a lecture in another language

Slice’s transcripts are made with AssemblyAI’s Universal-2 model, which its makers say covers 99 languages. They also publish how accurate it is in each, by word error rate: roughly, how many words in every hundred come out wrong. Their groups, as published on 4 November 2026:

High accuracy (about 1 word in 10 wrong, or better): English, Spanish, French, German, Indonesian, Italian, Japanese, Dutch, Polish, Portuguese, Russian, Swedish, Turkish, Ukrainian, Catalan.

Good accuracy (between 1 in 10 and 1 in 4 words wrong): Arabic, Azerbaijani, Bulgarian, Bosnian, Mandarin Chinese, Czech, Danish, Greek, Estonian, Finnish, Galician, Hebrew, Hindi, Croatian, Hungarian, Korean, Macedonian, Malay, Norwegian, Romanian, Slovak, Swiss German, Tagalog, Thai, Urdu, Vietnamese.

Moderate accuracy (between 1 in 4 and half the words wrong): Afrikaans, Belarusian, Welsh, Persian (Farsi), Armenian, Icelandic, Kazakh, Lithuanian, Latvian, Maori, Marathi, Slovenian, Swahili, Tamil.

Fair accuracy (more than half the words wrong): Amharic, Assamese, Bengali, Gujarati, Hausa, Javanese, Georgian, Khmer, Kannada, Luxembourgish, Lingala, Lao, Malayalam, Mongolian, Maltese, Burmese, Nepali, Occitan, Punjabi, Pashto, Sindhi, Shona, Somali, Serbian, Telugu, Tajik, Uzbek, Yoruba.

The rest of the 99 are transcribed but not in those groups. Source: AssemblyAI, Supported languages. These are averages over the company’s own tests. A clear recording close to the lecturer does better than its group; a noisy hall with an echo does worse.

The language is detected once, for the whole recording, and the transcript is written in it. That works well for a lecture given in one language from start to finish, which is most of them.

It works less well when the language changes:

• Language classes that move between English and the language being taught: the main language comes out well, the other poorly.
• A lecture with long quotations in another language: the quotation will be garbled. Check it against the source.
• An English lecture with a few terms in Latin, French or German: the terms will often be spelled as they sound. Check them against your slides.

If a recording is in two clear halves, one per language, cut it into two files and upload them separately, so each is detected on its own.

English comes out in American spelling. For a UK, Irish, Australian or New Zealand course, set your spell checker to your English before you quote from the transcript.

Accents are not a problem for most lectures. The rare failure is a lecture in strongly accented English detected as another language altogether: AssemblyAI’s own documentation says this can happen, and the transcript then comes out in that language. When the language is not English the result names it, so look at that and the first lines before you rely on it.

A transcript helps most when you are studying in a language that is not your first: you can read at your own pace, look up words, and put a paragraph you did not catch through a translator, with the timestamp to find it in the recording. Read the specialist terms against the slides or the reading list, since those are the words any transcript is most likely to get wrong.