French has more distance between sound and spelling than most European languages. Parler, parlé, parlée, parlés and parlez are pronounced the same. So are et and est, a and à, ces and ses, and the endings of ils mangent and il mange. A transcript therefore cannot be produced by ear alone: every one of those choices depends on the grammar of the sentence around it.
The speech model makes these choices from context, and in a clear, complete sentence it usually makes them correctly. The places where it slips are the same ones where French schoolchildren lose marks in a dictée: the agreement of a past participle with an object mentioned three clauses earlier, a plural that is only visible in writing, a rare proper name. If the transcript is going into a thesis appendix or onto a website, it is worth a careful read for agreements. For study notes and searchable minutes it is ready to use as it comes.
French from France, Belgium and Switzerland differs mostly in a handful of words, such as septante and nonante for seventy and ninety, and these are transcribed as spoken. Québec French is a bigger step: vowels are different, tu is added to questions, and informal speech is full of contractions and local vocabulary that standard spelling does not really cover. A university lecture or a radio interview from Montréal transcribes well. A relaxed kitchen conversation in joual will need more corrections, and we would rather say so than pretend otherwise.
French is also the working language of much of West and Central Africa and the Maghreb, often spoken alongside Wolof, Arabic or Lingala in the same sentence. The French part of such recordings is transcribed; short passages in another language may be rendered approximately or left out. Speakers who learned French as a second language are generally handled well as long as the recording itself is clear.
Spoken French drops things that written French keeps. The ne of negation disappears, il y a becomes y a, je ne sais pas becomes chais pas. The transcript stays close to what was said and uses standard spelling for it, so it reads as natural spoken French, not as a corrected essay. Verlan and recent slang are recognised when they are common and guessed at when they are not.
Office French borrows heavily from English: le call, un feedback, le planning, on va brainstormer. These are normally written as the English word. Figures deserve a second look in any language, and French makes them longer than most: quatre-vingt-dix-sept is one number, and a phone number read in pairs is easy to split wrongly. Check amounts, dates and reference numbers against the audio using the timestamp beside each segment.
Most French files uploaded here are between one and two hours long: a cours magistral, a research interview, a recorded committee meeting, a podcast episode. The free limit of 3 hours per file covers them in one piece, and guests get 10 such transcriptions. 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 PDF or Word export are part of the paid plans.
Recordings are processed by speech models on our own servers and are not sent to an outside speech API. For interviews collected under a research ethics protocol, or meetings covered by professional secrecy, that is often the first question to answer before any tool can be used.