Polish adds nine letters to the Latin alphabet: ą, ć, ę, ł, ń, ó, ś, ź and ż. They are not decoration. Łaska is grace and laska is a walking stick; sąd is a court and sad is an orchard; kąt is a corner and kat is an executioner. A transcript without diacritics has to be repaired word by word, which is why the text here is written in full Polish spelling from the first line.
Some Polish spelling cannot be heard at all. The pairs ó and u, rz and ż, ch and h sound the same in modern speech, so morze and może are told apart only by the sentence they stand in. The model chooses from context and is right in ordinary sentences. Rare words, and surnames such as Hojnacki or Chojnacki, Hałas or Chałas, are a matter of knowing the person, and those are worth checking against a list of participants.
A Polish name rarely appears twice in the same form. Pan Kowalski becomes pana Kowalskiego, panu Kowalskiemu, z panem Kowalskim; his wife is Kowalska, and the two of them are państwo Kowalscy. Foreign names are inflected as well, so a meeting about a client called Smith is full of Smitha and Smithowi. The transcript records these forms as they were spoken, which is correct Polish but means that searching for a name works best with its stem, not the full nominative.
Endings are short and unstressed, and in quick speech they blur: idę and idzie, tą and tę, robią and robiom. Standard spelling is used even when the speaker was careless, and most endings are recovered from grammar. When an ending is wrong, it is usually in a long sentence said quickly or in a noisy room. Numbers are a related case, because Polish inflects them too; dwóch, dwoma and dwiema are written as words or digits depending on context, so check figures that matter against the audio at the timestamp.
Recordings of Polish meetings contain a great deal of English with Polish endings: na callu, po deadlinie, zdeployować, zmerdżować, wrzucić na prod. The usual result is the English spelling with the Polish ending attached, but there is no single accepted way to write many of these and the transcript is not always consistent between one occurrence and the next. Product names, internal project names and abbreviations are the other regular source of corrections in business recordings.
Polish is unusually uniform across the country, and a speaker from Gdańsk, Kraków or Lublin is transcribed equally well. The exceptions are real dialects and regional languages. Silesian, Kashubian and the highland speech of Podhale differ from standard Polish in sounds and vocabulary; a recording in them will be written as the nearest standard Polish, with gaps. Poles living abroad who mix in English or German phrases are transcribed well in the Polish parts, less reliably where the language switches mid-sentence.
The usual Polish upload is one to two hours long: a recorded company meeting, a university lecture, an interview for a thesis or an article, a hearing or consultation recorded on a phone, a long podcast from YouTube. Guests get 10 transcriptions of up to 3 hours each. The free allowance is 5 hours in total across those 10 recordings; a free account keeps the full transcripts, with a history and an editor. AI chat with the transcript, automatic summaries and export to PDF or Word are part of the paid plans.
The speech models run on our own servers, and recordings are not passed to an external speech API. For law firms, medical practices and anyone who has to explain to a client where a recording was processed, that is a short and simple answer.