Verbatim versus cleaned dictation
When you want every word you said, and when you want the sentence you meant. Two modes, and how to tell which one a piece of writing needs.
Voicecape has two modes, Exact Clean and Verbatim, and the difference between them is not quality. The same speech model produces the same transcript either way. What changes afterwards is whether a short list of rules removes hesitation sounds, doubled function words and a few trailing phrases before the text lands in your document.
That makes the choice a question about the destination rather than about the dictation. Some of what you produce will be quoted, cited or checked against a recording. Some will be read once and acted on. Those two want opposite things, and choosing by which one you are making is more reliable than choosing by which output looks better in a demo.
It is worth saying up front what Verbatim is and is not, because one filter runs in both modes and it would be dishonest to describe Verbatim as a byte-for-byte copy of the audio.
Text you are going to quote
Use Verbatim when the words themselves are the artifact.
Interview notes are the clearest case. If you are transcribing what someone told you and any of it may end up in quotation marks with their name beside it, the hesitations are part of the record. A pause before a number, a word started and abandoned, a repetition -- those carry information about how confident the speaker was, and a cleaner that tidies them is editing a source.
Legal and medical dictation is the same argument with higher stakes, since in both the phrasing is often the substance: a qualifier, a repeated instruction, a correction spoken aloud. Language practice is the case that decides itself -- if you are dictating in a language you are learning, the disfluencies are exactly what you were trying to see, and cleaned dictation hides the evidence you asked for.
The general rule: if the text will later be presented as what someone said, do not let anything decide on your behalf which of those words mattered.
Text you are going to send
Use Exact Clean, the default, when what you want is the sentence you meant.
Email, chat, issue comments, commit messages, documentation, notes to yourself, first drafts of anything. In all of these the reader has no interest in your hesitations, and neither do you when you reread it next week. Speaking a paragraph is faster than typing it, and the tidying afterwards is the only reason people go back to the keyboard. Removing three ums is the difference between dictation that saves time and dictation that moves the work rather than removing it.
Cleanup never touches identifiers, paths, numbers or repeated digits, so dictating a variable name or an error code is safe in either mode -- dictation for developers goes into that. What makes Exact Clean usable for text you will send is that its edits are a fixed, short, enumerable list rather than a judgement about your phrasing. The full list is written out word by word, including what it leaves alone -- self-corrections above all, since deciding that send it Friday, no, Thursday means Thursday takes prosody that text does not carry.
The one thing both modes do
Verbatim is not a raw dump of the recogniser’s output, and it should not be described as one. A single filter runs before the mode branch, in both modes: recogniser artifacts are dropped.
Speech models trained on large amounts of subtitled video learn to emit annotations as well as speech. Given silence, background noise or music, they will sometimes produce a line like (Music), [BLANK_AUDIO], (Applause) or (speaking in foreign language) -- text describing the audio rather than transcribing it. Nobody said those words, and inserting them would break the one thing the product promises, which is that only what you said arrives.
So Verbatim means untouched speech, not untouched output. It is a promise about your words, not a promise to pass along things the recogniser invented.
What counts as an artifact
The test is deliberately strict, because the cost of getting it wrong is deleting something a person said.
A line goes only when the whole line, after removing exactly one enclosing pair of brackets -- round, square, curly, or a pair of music notes -- matches an entry in that language’s list exactly. Substring matching is forbidden. Several annotations stacked on one line are peeled one at a time, since real output has been observed with two in a row, and the line goes only if every piece is an annotation.
The lists are per language and short. English has entries like blank_audio, inaudible, music, applause, laughter, silence, speaking in foreign language and end; Korean has 음악, 박수, 웃음, 무음, 끝 and a few subtitle credits. Each list is limited to entries observed in real output or documented in the recogniser’s own token set, because a speculative entry deletes real speech.
- The music was loud. survives -- music is a word people say, and the line matches nothing.
- I said (music) out loud. survives -- the bracketed part is not the whole line.
- (the short one) is fine. survives -- brackets prove nothing; the contents have to be a known annotation.
- 이걸로 끝입니다. survives, while a line consisting only of (끝) does not.
The subtitle sentences with no brackets
Japanese and Chinese have a harder case. In silent passages these models will sometimes produce a complete, ordinary-looking sentence borrowed from subtitle data -- a sign-off like ご視聴ありがとうございました, or a channel plug like 请订阅我们的频道. Nothing on the surface marks them as non-speech.
So the rule for those carries an extra condition: such a line is removed only when at least one other line of the transcript survives. If the boilerplate sentence is the entire transcript, it is kept -- you may simply have said it.
That condition exists because of an asymmetry that shapes the whole engine. A wrongly deleted line is a loss you cannot see; a wrongly kept line is one you can read and delete. When the two cannot be told apart from the text alone, it keeps.
Why this is a filter and not a recogniser setting
There is a decoder flag in whisper that suppresses non-speech tokens, and it is the obvious place to solve this. It does not work here, for two reasons.
First, it operates on symbol tokens -- characters like @ / : _ and the double quote -- by pushing their probabilities down. Those are exactly the characters that make up email addresses, URLs, file paths, code identifiers and quoted strings. Turning it on to block one hallucination costs you all of those, at the decoder, with no way to recover them.
Second, it would not catch the problem anyway. Phrases like speaking in foreign language, 음악 and 끝 are made of perfectly ordinary word tokens, and the flag never sees them. A filter after the fact can be narrow, can be read in a text file, and can be argued with.
What else is shared
Both modes normalise whitespace: runs of spaces and tabs collapse to one, edges are trimmed, runs of blank lines are capped at two. Line breaks survive, so the structure you spoke is preserved.
Those paragraph breaks are measured rather than guessed. The recogniser reports the start and end of every segment, and a gap longer than about a second and a half becomes a blank line -- long enough that hesitations inside a sentence do not split it, short enough that three spoken paragraphs do not arrive as one block. The behaviour is identical in both modes, because it reflects how you spoke rather than what you said.
All of it happens on your Mac; neither mode sends anything anywhere. That is worth knowing if the reason you reached for Verbatim is that the recording was sensitive to begin with -- see on-device versus cloud speech recognition.
A working rule
If the text will be attributed, cited or checked against a recording, dictate in Verbatim. If it will be read and acted on, dictate in Exact Clean. The mode is a setting, so switching for one interview and switching back afterwards takes a couple of seconds.
When you are unsure, Verbatim is the safer default for that particular piece of text: the tidying you did not get is a keystroke away, and the words you did not lose cannot be recovered any other way.