Verbatim vs clean transcript
A verbatim transcript captures everything: every "um," every false start, every repeated word, exactly as spoken. A clean transcript smooths that out into readable sentences, the way a person would write down what someone meant to say. A model learning to recognize real speech usually needs verbatim. A person skimming a meeting summary usually wants clean.
We ask which one you need as part of the brief, since the same audio produces two different transcripts depending on the answer.
Timestamps and speaker labels
Word-level or segment-level timestamps let a model learn exactly when each word was spoken, which matters for anything that has to sync audio and text. Speaker labels, who said what, matter the moment there's more than one voice in the recording, whether that's a call center conversation or a meeting.
Timestamps and speaker labels are part of the standard delivery, not an extra we charge separately for.
Noise, fillers and overlapping speech
Real audio has background noise, people talking over each other, and words that get swallowed or unclear. A transcript for AI training data usually tags these rather than guessing or leaving them out, since a model needs to learn what uncertain audio actually sounds like, not a cleaned-up version of it.
Unclear audio gets tagged, not guessed at, and overlapping speech is marked rather than merged into one line.
Formatting for a model vs a reader
A model doesn't care about paragraph breaks or punctuation for readability, it cares about consistent, structured formatting it can parse the same way every time. A person reading a transcript wants punctuation, paragraphs and a style that reads naturally. Mixing the two goals in one transcript usually serves neither well.
We agree the format upfront, structured for a model or readable for a person, rather than defaulting to one and hoping it fits.
How to decide which one you need
Ask what happens to the transcript next. If it trains or evaluates a model, lean verbatim, timestamped and structured. If a person reads it to understand a conversation, lean clean and readable. Some projects genuinely need both: a verbatim version for training and a clean version for a human reviewer.
Tell us what the transcript feeds into, and we'll tell you which format actually fits, rather than defaulting to one for every job.
Warning signs
Be cautious of a vendor who doesn't ask which kind of transcript you need, delivers clean text with no option for verbatim, or can't explain how they handle unclear audio and overlapping speech.