Languages

We transcribe 25 languages. Some come out nearly clean, some need a read through. Here is the measured error rate for each, so you know before you pay.

The number is how many words in a hundred the speech model gets wrong on a standard test of clear, read speech (FLEURS). Real recordings with noise, crosstalk or people talking over each other do worse; a good microphone and one speaker at a time do about this well. The language is detected from the audio itself, so a recording that switches language part way is still transcribed.

Strong

Clean enough to read and quote after a quick look. Expect to correct names.

Language Words wrong in 100
Italian 3.0
Spanish 3.5
Portuguese 4.8
English 4.8
German 5.0
French 5.2
Russian 5.5
Ukrainian 6.8
Polish 7.3
Dutch 7.5

Good

Reads well. Expect to fix names and a word here and there.

Language Words wrong in 100
Slovak 8.8
Czech 11.0
Romanian 12.4
Croatian 12.5
Bulgarian 12.6
Finnish 13.2
Swedish 15.1
Hungarian 15.7

Usable, check it

Good for searching a recording and for a first draft. Read it against the audio before quoting.

Language Words wrong in 100
Estonian 17.7
Danish 18.4
Lithuanian 20.4
Maltese 20.5
Greek 20.7
Latvian 22.8
Slovenian 24.0

Not yet

We do not transcribe these. A recording in one of them comes back wrong or empty, so please do not pay for one: Norwegian, Arabic, Chinese (Mandarin), Cantonese, Japanese, Korean, Hindi, Turkish, Hebrew, Indonesian, Vietnamese, Thai, Persian, Urdu, Bengali, Tamil, Serbian, Bosnian, Icelandic, Catalan, Irish, Welsh, Afrikaans, Swahili, Malay, Filipino. The same goes for any language not in the lists above.

We also do not translate (the text comes back in the language that was spoken), and we do not label who is speaking.

The model is NVIDIA Parakeet TDT 0.6B v3, an open speech model. The figures are from its model card, read on 2026-11-17.

Drop your files