How to Transcribe Non-English Audio
Transcribing a recording that is not in English: setting the right language, what accuracy looks like across languages, and what still does not work.
The Scriptorium team ·
Upload the recording and set its language explicitly rather than leaving it on auto-detect, if you know what the language is — the process supports 33 languages plus automatic detection, and picking the right one directly is more reliable than letting detection guess from a few opening seconds.
Setting the language
Auto-detect works by listening to the first few seconds of audio, which is enough most of the time but can commit to the wrong guess if a recording opens with music, silence, or a brief line in a different language than the rest of the recording. Setting the language explicitly whenever you know it removes that failure mode entirely.
Accuracy is not the same across languages
This is worth saying plainly rather than glossing over: speech recognition quality varies by language, and it is not evenly distributed. Languages with more available training data and simpler phonetic-to-text mapping tend to transcribe more reliably; languages with less data, heavy dialectal variation, or no standard writing system for casual speech tend to do worse. Nobody offering transcription, including this one, should claim uniform accuracy across every supported language, and none of the figures on this site do.
Nobody should quote you a single accuracy figure, including us. Automatic transcription of one person speaking clearly into a close microphone is very good — good enough that correcting it is much faster than typing it. The same system on a recording with two people talking over each other, a noisy room, a strong accent or a lot of specialist vocabulary is meaningfully worse.
The failure mode worth knowing about is that it does not degrade into obvious nonsense. It produces fluent, confident, plausible sentences that are wrong, which is harder to catch than gibberish would be. That is why you check a transcript against the audio rather than reading it on its own.
A recording with an accent, or informal speech
Strong regional accents and informal or dialectal speech are harder than the standard, clearly-enunciated version of a language that most training data reflects — this is true within English as well, not just across languages. It is not a reason to avoid transcribing the recording, but it is a reason to check the result against the audio rather than trusting it outright, especially for names and regionally specific vocabulary.
A recording that switches languages
Covered in more depth in its own article — the short version is that a recording with more than one language in it is a genuinely harder case, and choosing whichever language dominates the recording, then expecting to correct the other language's segments by hand, is the realistic approach rather than a fully automatic one.
What is not supported
A language outside the 33 listed on audio to text and video to text is not supported at all, full stop — there is no partial or degraded mode for an unlisted language, and claiming otherwise would not be honest.