Verbatim transcription: the exact words first, the summary built from them

Verbatim transcription is a written record of the words people actually said, rather than a retelling of them. In Dalm, a Mac app for dictation and meeting notes, the transcript of a meeting works this way: each line carries the words as they were recognised, with its time and a voice number. Speech is recognised on your Mac and no audio is sent anywhere; the transcript itself is synced to our servers automatically. A language model then writes the summary and documents from that text, and each point of the summary holds a word-for-word quote, so the summary can be checked against what was said.

Where the exact words matter more than a summary

A summary tells you what was decided. When the wording itself is what you need, it is the wrong record:

• a quote for an article or a report, where a person’s own words go into print under their name — more for journalists;
• an interview for a study, where how someone put a thing matters as much as what they meant;
• an agreement someone may later remember differently: who offered what, and in which words.

In each of these a retelling puts other words in the speaker’s mouth, the note-taker’s or a model’s. A verbatim transcript keeps the speaker’s own. Minutes are a third kind of record again: they set down what a meeting decided, not how it was said, and how to write them is in the guide to meeting minutes.

Verbatim, clean verbatim and a summary: three kinds of record

The words verbatim and clean verbatim come from transcription services, which sell them as two kinds of transcript:

• Verbatim: every word as it was spoken, filler words such as um and uh, repeated words and false starts included; some services also mark laughter or two people talking at once.
• Clean verbatim: the same words with the fillers and false starts taken out, and nothing else changed.
• A summary: shorter, in someone else’s words, about what was said rather than how it was said.

The first two keep the speaker’s wording; the third replaces it. The Dalm transcript belongs with the first two: no line is retold or shortened. The first two differ only in the fillers and false starts. Dalm takes nothing out itself: fillers, repeats and false starts stay as the speech model writes them down, and in our English test um and uh stayed in. There is no mode that cleans them up, so what you get sits closer to verbatim than to clean verbatim.

How Dalm keeps the words: the transcript comes first

After a meeting, speech is recognised on your Mac and split by voice, and the app puts the transcript together line by line, each line with its time and its voice. No language model writes or rewrites the words. Punctuation and capital letters come from speech recognition itself, and if you keep a dictionary, a term that was heard wrongly is swapped for the spelling you set. No audio is sent anywhere; the transcript, with its times and voice numbers, is synced to our servers automatically, and no setting turns that off.

You can correct a line: Edit, then ⌘↩ to save. Show original brings back what the line said before your edit, so a correction never hides what was recognised. A line you correct reaches your other Macs, and later questions and documents use the corrected text; a summary already written is not rebuilt. How the times work is on transcription with timestamps, and how voices are told apart on speaker diarization.

From the transcript to the summary, and back to the words

The summary and the documents are written afterwards by a language model on our servers, from the text of the transcript; the audio is not sent to it or anywhere else. The model is told not to invent: where something was not said, the summary leaves it blank or marks it as not said instead of filling the gap with a guess. Each point of the meeting summary carries a quote in the speaker’s own words. Every quote shows you where it was said — open it and the transcript opens at that place.

That is the order the app keeps: the transcript is the record, and everything shorter is built from it and can be checked against it. A summary you cannot check is a retelling you have to trust; one that points back to the words is a shortcut through them.

People who transcribe, or an app: what each can promise

A transcription service that works with people sends your recording to a person, who listens and types. That is the way to get fillers kept or taken out to order, unclear passages marked, and, where a service offers it, a transcript signed for use as a record. What it costs is that the recording leaves your hands, and you wait while a person works through it.

An app does the listening itself, and nobody else hears the conversation. In Dalm that happens on your Mac, and the text then goes to our servers, as described above. The trade runs the other way: nobody checks the result but you, so the transcript is a working record, not a certified one.

What a transcript without the audio cannot do

• No sound to check a word against. The audio is deleted once the text exists. At the foot of every transcript the app says: “Audio was deleted once the transcript was written.” If you may need to hear a moment again or play it to someone, keep a recording made another way; the ways that leave you a file are compared on meeting recorders.
• Not a certified record. A transcript used as evidence or in a formal proceeding follows rules set by whoever will rely on it. Dalm does not claim to meet them; whether a transcript will do is a question for them or for a lawyer.
• Recognition can mishear. A name, a term or a quiet voice may come out wrong. Edit fixes the line, and the dictionary fixes a term for the recordings that follow; more on names and your own terms.
• The text leaves your Mac. The transcript is synced to our servers automatically, and a language model reads it there when the summary is written. What goes where is set out on privacy.

Free, Pro or Team: after the first verbatim transcript

The transcript, the summary written from it and the quotes inside it all come with the Free plan, which is a plan and not a trial. You need a Mac with Apple Silicon · macOS 15.0 or later and an account, and while the app is not public, you get in through the waitlist. What the paid plans add is on the pricing page. In a team the dictionary is shared, so a name fixed once is spelled the same way in the transcripts that follow, for everyone: more on working as a team.