In Czech the marks above letters are not decoration. "Byt" is a flat and "být" is to be; "pravá" and "práva", "zeď" and "zed" are different things. A transcript without diacritics, or with them guessed afterwards, has to be repaired by hand. Here the text is written in full Czech orthography from the start, including ú at the beginning of a word and ů inside it.
Vowel length is where spoken Czech is least careful. People shorten "říkám" to something like "řikam" and "prosím" to "prosim", and the model normally restores the standard length. The pairs i and y, and mě and mně, sound the same and are chosen by grammar and context; the result is right far more often than not, but agreement in past-tense plurals ("ženy šly", "muži šli") is the place where a slip is most likely.
Almost nobody speaks the way Czech is written. In Bohemia people say "dobrej", "velký okna", "bysme", "vokno"; the standard forms are "dobrý", "velká okna", "bychom", "okno". The transcript does not settle this the same way every time. Common colloquial endings are sometimes kept as spoken and sometimes turned into the standard form, and prothetic v- is usually dropped.
What to do about it depends on the purpose. For minutes and news quotations, a quick edit towards standard Czech is normal practice. For sociological or linguistic interviews, where the spoken form is the data, the automatic text is a first draft that needs the colloquial forms put back. Either way, the timestamp on each segment takes you to the exact place in the audio.
Czech inflects names, which makes them harder than in most languages. "S panem Novákem", "od Dvořákové", "u Jirky" are all forms the model has to recognise, and it does for common surnames. Rare surnames, foreign names with Czech endings ("s Johnem Smithem") and company names in the locative are the least reliable items in a transcript; compare them with an attendance list. Numbers appear as digits and should be verified for dates and sums.
Standard Czech and common Bohemian speech are recognised best. Moravian speech, with its shorter vowels and words like "su" and "chcu", is usually readable. Strong Silesian or Lach dialect near Ostrava and broad Haná speech bring more errors. Slovak is a separate language to the model: a Slovak colleague in a mostly Czech meeting may be written in Slovak, in Czech, or in a mixture, and those passages are worth a closer look.
Czech files uploaded here are mostly interviews for bachelor's and master's theses, journalistic interviews, company meetings, municipal council sessions, lectures and podcasts. Singing over music transcribes poorly, so this is for speech. Guests get 10 transcriptions of up to 3 hours each and read a free preview; the free allowance is 5 hours in total, and a free account keeps the full transcripts. AI chat with the recording, automatic summaries and PDF or Word export are in the paid plans, and subtitles can be translated into other languages.
Audio is processed on our own servers with our own speech models and is not handed to a third-party speech API. For recordings containing personal data, that keeps the list of processors under GDPR short.