How Dalm works: what stays on your Mac, what doesn't
Dictation, video calls and in-person meetings. Audio is recognised on your device by default; the text syncs so documents and answers can be built from it.
By default, speech recognition happens on your Mac, and so does the work of separating one voice from another. The audio file is deleted once the text is ready — there is no setting to keep it; if processing fails, it stays on your Mac — and by default the sound itself goes nowhere. The text leaves: the transcript syncs to the server automatically, without a separate action from you, and that is where documents and answers about a conversation are assembled. One line, stated once: by default the audio stays, and the text always leaves. The table below says which side of that line every kind of data falls on.
You can choose server recognition in Settings, separately for meetings and for dictation: then the audio is sent to Dalm and our speech recognition providers and deleted once the text is ready. Either way, the text is synced to our servers.
The short version
- Speech recognition runs on your Mac by default, in both accuracy models.
- The audio file is deleted once the text is ready. There is no setting to keep it.
- The transcript syncs to the server automatically, with no separate action from you and no setting that stops it. What syncs is the text: utterances with their timestamps and voice numbers.
- Anonymous usage statistics have a switch, off by default, and in the current version the app sends no statistics whichever way it is set.
- Deleting your account removes its data from the server: your dictations, and every workspace where you were the only member, with its meetings, notes, chats and tags. Deleted items wait 30 days before they are erased, and copies in our backups can stay for up to 30 days longer.
What happens when you dictate, call or meet?
Dictation. Press the global hotkey in any application with a text field, talk, and press it again. The text is put into whatever field the cursor is in, so there is no window to switch to. Recognition happens on the machine, which is why this works the same with the network off.

A video call. Start recording and the app captures the system audio and your microphone as two separate tracks. Nothing joins the call: there is no participant to admit and nothing appears in the attendee list, on Zoom, Meet, Teams or Telegram alike.

Two tracks separate you from everyone else reliably, because that separation is mechanical: your microphone is one track, everything your Mac plays is the other. Telling apart the individual people inside that second track is a different thing and a weaker one — the model works it out from context, and it can be wrong about who said which line.
What it does not do is settle a name on its own. Until you answer about a voice, its place in the transcript holds the model’s guess or a numbered speaker, and in the naming step a guess may be marked “Suggested by Dalm — check before you agree.” In the transcript and the summary a guess looks like any other name until you confirm or change it, so check the names before you send a summary on. A name becomes a name when you confirm it.
A meeting in the room. The same path through the microphone. No calendar invitation, no link, no second device — you start the recording and the indicator shows it is on until you stop it.
Timings are deliberately absent from this page: a number is published here only together with the method used to measure it, and until that method exists the honest description is in words.

Where does each piece of data live?
| What | Where it happens or lives | Does it leave |
|---|---|---|
| Audio capture, microphone and system audio | On the device | No |
| Speech recognition, both accuracy models | On the device | No |
| The audio file | Deleted once the text is ready; there is no setting to keep it. For a file you import, Dalm makes a working copy, deletes it once the text is ready, and never changes the original | No |
| Transcript | Synced to the server automatically: utterances with their timestamps and voice numbers | Yes, automatically. There is no separate action and no way to switch it off |
| Files you add: materials attached to a meeting, files attached to a question in chat | On the server, as separate files in private storage that opens only through signed links. An attachment never sent with a question is removed after one day; a deleted file can be restored for 30 days, then it is erased. No audio is stored there | Yes, when you add them |
| Tags | Synced to the server; in a team, shared across the organisation | Yes, automatically |
| A summary, a document or a question to the assistant | Nothing new reaches our server: the text has been there since the recording. What happens is the other direction — the text goes to third-party AI model providers for the first time. The summary is written automatically after each meeting; documents and answers when you ask | Yes, to third-party AI model providers |
| Sharing a document | “Share” hands the document’s text, in Markdown, to the macOS share menu — mail, Messages, AirDrop. Sharing a meeting by email sends a link to its summary, and you can revoke it | Only to where you send it |
| Calendar | A connected Google Calendar goes through the server, which keeps its access token encrypted, and meetings you schedule in Dalm are written into the calendar you choose | Scheduled meetings are written to your calendar |
| Anonymous usage statistics | A switch at setup and in settings, off by default. In the current version no statistics are sent, whichever way it is set | No |
| What the model receives | On processing: the transcript with its utterances, your own notes, and the names you gave the voices. On a question in chat: the same, plus the conversation so far and your question. On a document view: the same, plus the assembled result | Yes, to third-party AI model providers |
| Automatic emailing of summaries to attendees | We do not do this in any form | Never |
With the network off, these keep working: dictation, recording, speech recognition. What waits for a connection is the server side — summaries, documents and questions about a meeting, and the Google Calendar connection.
Two things this table does not yet answer, and both belong here rather than in a policy nobody opens. Transcripts stay on the server until you delete the meeting: there is no retention period. The database and files are backed up, and a deleted item can stay in the backups for up to 30 days. A setting for keeping the transcript changes what you can do with a meeting afterwards: with it turned off, the summary and the documents stay, but a line no longer links to the moment it was said, and you cannot ask questions about the quotes.
What do you control?
A recording you can see and stop. While a meeting records, the menu bar icon carries a dot and the recording screen shows a timer and that recording is on. Stop ends it; Cancel throws the recording away at once. A keyboard shortcut starts and stops a meeting from any app.
A note on consent before your first meeting. Before the first recording the app shows its note on recording and consent: some places require everyone’s agreement, and the person recording is responsible for it. It suggests saying out loud that you are recording. The app itself tells nobody — it has no way to reach the other people on the call, and nothing is posted to the call or its chat.
Deletion that reaches the server. Deleting your account removes your dictations and every workspace where you were the only member — its meetings, notes, chats and tags. Deleted items wait 30 days before they are erased, and copies in our backups can stay for up to 30 days longer. It does not cancel a paid subscription; that is a separate step. Signing out wipes the copy on that Mac: the local database and the search index. If you delete your account, meetings you recorded in a team workspace stay with the team.
Three things stay on this Mac and are never part of the sync. The audio itself. The timing of each word. Your application settings. None of the three is on the list of what travels, and none of them is something you have to switch off to keep.
What should you ask any vendor?
Six questions worth asking before you trust a tool with your conversations. They are not about us, and the answers differ across the category.
- Where does speech recognition happen — on the machine or on a server?
- Is the audio kept after transcription, and for how long?
- Are models trained on customer data, and can that be switched off?
- What exactly goes into telemetry: events only, or content?
- Who can open a shared link, and can it be revoked?
- What happens to the data when the subscription ends?
Where to go next
- Privacy in plain words — the same boundary, without the table.
- Pricing — what the free plan covers and what Pro adds.
- What an AI notetaker is — the category, explained from scratch.
- Recording consent laws — what the rules are where you work.
- Local against cloud — the trade-off itself, in more depth.
- Meeting notes without a bot — how capture works when nothing joins the call.
Questions
- Where is my meeting data stored?
- In two places, split by kind. By default, audio is not sent anywhere: it is recognised on your Mac, and the audio file is deleted once the text is ready; if processing fails, it stays on your Mac. The text — transcripts, notes, documents, tags and chats with the model — is stored in your account on the Dalm server, which is why it is on every Mac you sign in to. The search index and the timing of each word stay on the Mac. The table on this page lists every kind of data, one row each.
- How does speech recognition work here?
- Both accuracy models run on your Mac by default. The app captures the microphone and the system audio as two separate tracks, turns them into text locally, and then deletes the audio file. Nothing about that step needs a network connection, which is why dictation and recording keep working with the network off. Assembling documents and answering questions about a conversation is a separate step and a separate place — the server. If you choose server recognition in Settings, the audio is sent to Dalm and our speech recognition providers and deleted once the text is ready.
- Is the audio deleted after the recording is transcribed?
- Yes. The audio file is removed once the text is ready, and there is no setting to keep it; if processing fails, it stays on your Mac. For a file you import, Dalm makes a working copy, deletes it once the text is ready, and never changes the original. By default, audio is not sent to the server: what travels is text, and only the kinds of text named in the table on this page.
- What ends up in telemetry?
- Crash reports go to Sentry (EU, kept 30 days), and there is the “Report a Problem…” report you send yourself. No usage statistics are sent: the switch at setup and in settings is off by default and only stores your choice. The text of your records, the audio and the titles of your meetings are not part of it. The table on this page gives it its own row rather than folding it into something vaguer.
- What happens to my data if I delete my account?
- Deleting the account removes its data from the server: your dictations, and every workspace where you were the only member, with its meetings, notes, chats and tags. Deleted items wait 30 days before they are erased, and copies in our backups can stay for up to 30 days longer. It does not cancel a paid subscription; that is a separate step. What lives only on your Mac — the search index and the timing of each word — is wiped when you sign out of the app on that Mac.
- Can it work without a network?
- Dictation, recording and speech recognition run with the network off. What does need a connection is the server side: summaries, documents and questions about a meeting, and the Google Calendar connection. Those wait until you are online again rather than failing quietly.
- What does the language model receive, and when?
- Nothing new reaches our server — the text has been there since the recording. The summary is written by itself after each meeting; documents and answers when you ask. What is new goes the other way: this is the moment the text first leaves for third-party AI model providers. Processing sends the transcript with its utterances, any notes you wrote, and the names you gave the voices. A question in chat sends the same, plus the conversation so far and your question. Opening a document view sends the same, plus the assembled result. No audio in any of the three.
- Can I turn the syncing off, or use it without an account?
- No to both, and it is better to know that now than after installing. An account is required: without signing in there is no window and no dictation, so there is no signed-out mode to fall back on. Syncing cannot be switched off either — the transcript goes to the server automatically and no setting stops it. What stays on the machine stays there whatever you do, but what syncs, syncs.
- Does it join my call as a participant?
- No. The app records the audio your Mac already plays and hears, so there is nothing to admit and nothing appears in the attendee list. A video call on Zoom, Meet, Teams or Telegram is captured the same way a phone call on speaker or a meeting in the room is captured, and nobody has to let software into the conversation.