A great deal of informal Greek online is typed in Latin letters, and anyone who has had to turn "kalimera ti kaneis" back into καλημέρα, τι κάνεις; knows how slow that is. The transcript here is written in the Greek alphabet from the start: η, ι, υ, ει and οι are chosen by spelling, not by sound, even though all five are pronounced the same, and ο and ω are kept apart the same way.
The stress accent, the τόνος, is placed on words of more than one syllable, which matters more than it looks: πότε and ποτέ, γέρος and γερός, νόμος and νομός are different words separated only by that mark. The Greek question mark, which looks like a semicolon, is used where a question was asked. The old polytonic system with breathings is not produced.
Because several letters share one sound, homophones are the typical error. The endings -ει and -η, -ετε and -εται, or the article της against the pronoun τις are decided by grammar, and in fast or mumbled speech the model sometimes picks the wrong one. These are quick to fix, since the word is right and only the spelling of its ending is off.
Final ν is another place to check. Greek speakers drop it or keep it by habit (τον/το, την/τη, δεν/δε), and the transcript tends to follow what was actually said, which may not match the rule a proofreader would apply. Text meant for publication benefits from a pass through a spell checker.
Standard Modern Greek as spoken in Athens and Thessaloniki, and on television and radio, is recognised best. Cypriot Greek differs in sounds and vocabulary, and the transcript will often normalise it towards standard forms or miss dialect words altogether, so expect more correction for recordings from Cyprus. The same applies to strong Cretan or Pontic speech.
Business and technical talk in Greek is full of English: meeting, deadline, marketing. These may appear in Latin letters or in Greek spelling (μίτινγκ), sometimes both in one file. Surnames deserve a check too. Παπαδόπουλος is safe, but less common family names and place names are spelled by ear, and a participant list settles them in a minute.
Greek files uploaded here are mostly university lectures, court and council recordings sent for απομαγνητοφώνηση, research interviews, podcasts and YouTube talk shows. Songs are a different matter: sung lyrics over bouzouki or a full band transcribe poorly, so this is a tool for speech. Guests get 10 transcriptions of up to 3 hours each, with 5 hours in total across those recordings. Without an account a free preview of the text opens; a free account keeps the full transcript. AI chat with the recording, summaries and PDF or Word export come with the paid plans.
The audio is processed on servers we run, with our own speech models, and is not forwarded to a third-party speech API. For interviews with named sources and internal company meetings that is usually the first question asked before anything is uploaded.