Narada
Download for Mac
Answer

Which meeting notes apps keep the audio, and which delete it

A summary is a lossy compression of a conversation. Sooner or later you will need the sentence rather than the gist, and by then the audio either exists or it does not.

By Ivan Stepanok, developer of Narada · Updated

Short answer
  • Granola keeps no audio, and says so: “It does not record or save audio or video at any point during the call, so there’s no way to access audio from your meetings.”
  • Fireflies' bot-free desktop mode keeps none either — “System audio mode does not save separate audio or video files” — and Jamie deletes it after transcribing: “we delete all audio files permanently.”
  • The tools that do keep audio mostly keep it in their cloud under their retention rules. tl;dv’s privacy policy sets recording retention at “Free user: 3 months Paying user: until account deletion”.
  • Narada writes one .m4a per meeting into a folder you chose, next to the transcript — 48 kHz mono AAC at 128 kbps, roughly 58 MB an hour.
  • Keeping the file is not free. It costs disk, it is unencrypted on purpose, and a recording of other people is a responsibility you took on.

Most AI notetakers give you text and nothing else. Some never write the audio to disk, some upload it and delete it within minutes of transcribing, and the ones that keep it mostly keep it in their own cloud under a retention policy you do not control.

There is a decent argument for that — audio you never store is audio you can never leak — with a consequence people discover late. The day somebody disputes what was agreed, the recording is either on your disk or it is gone.

Who keeps it, quoted from their own docs

Every cell below comes from the vendor’s own documentation or pricing page, read 4 August 2026.

ProductIs there an audio file afterwards?Where it lives
GranolaNo — “there’s no way to access audio from your meetings”Nowhere; streamed to a transcription provider and dropped
JamieNo — “we delete all audio files permanently”Uploaded to Frankfurt, processed on Modal GPUs, then deleted
Fireflies, bot-free desktop modeNo — “System audio mode does not save separate audio or video files”Transcript and summary only
Fireflies, bot modeYesFireflies' cloud, US (AWS and GCP)
OtterYes — mp3 is an export format on every tierOtter’s cloud; custom retention is Enterprise-only, via an account manager
tl;dvYesTheir library — “Free user: 3 months Paying user: until account deletion”
CirclebackYes — the Individual plan lists “Meeting video and audio playback”Their cloud, US, retention settings are a Team-tier feature
FathomYes — “Call downloads and clips (unlimited)" on every planFathom’s cloud, US; its bot-free capture is footnoted “Beta feature for Mac”
Teams intelligent recapPartlyRecording in OneDrive/SharePoint under tenant policy; audio recaps “stored for 60 days” in the user’s OneDrive
NaradaYesrecording_HH-MM-SS.m4a in your own folder, beside the transcript

Two rows deserve reading twice. Granola’s is the clearest statement anyone in this category has published, and the audio genuinely never touches their storage. The Fireflies pair is the one that surprises people: picking the bot-free mode to keep an extra participant out of the call also opts you out of having a media file, so the only route to a recording there is to let the bot join. Longer comparison of the bot-free mechanisms.

Four reasons the file matters

A disputed sentence. “You said you’d cover the migration” is settled by the recording and not settled by an AI paraphrase of it.

Speech models keep getting better. Whatever transcribed your 2026 meetings will look crude in 2029. Keep the audio and you can re-run the better model over the old archive; keep only the transcript and its errors are permanent. That bites hardest on accented speech, cross-talk and languages today’s model handles poorly.

Your retention policy should be yours. A file on your disk expires when you delete it. A recording in a vendor’s account expires when their default says so — three months on tl;dv’s free tier — or when an admin changes a setting, or when you stop paying.

Feeding it to something else. Better diarisation, a translation tool, an editor to clip 40 seconds for a colleague — all of it needs a file, and the apps that keep no audio cannot even accept one: Granola’s docs note it “does not support importing or uploading pre-recorded audio files (such as MP3s) for transcription.”

“Kept” is not the same as “yours”

Several products above do keep audio, and if all you want is to re-listen, they are fine. The distinction is custody.

Microsoft’s intelligent recap documents the dependency plainly: “If recording is turned off, users experience recap without the recording, speakers, topics, and chapters.” Whether a recording exists is an admin’s decision, and the files land in OneDrive or SharePoint under the tenant’s retention — recoverable by the organisation, not only by you.

Otter is the friendlier version of the same shape. The mp3 is there on every plan, which is more than most of this market offers, but it sits in Otter’s account until you export it, and setting a retention period is an Enterprise feature arranged through an account manager rather than a switch in settings.

Where Narada puts it

Narada records two streams — your microphone and your Mac’s system output, via the macOS system-audio permission and no virtual audio driver — and writes a folder per meeting:

~/Documents/Narada Meetings/2026-08-04/Quarterly review/
  recording_14-30-02.m4a           # 48 kHz mono AAC, 128 kbps
  recording_14-30-02.txt           # transcript, [M:SS.t] Speaker: text
  recording_14-30-02_summary.md
  recording_14-30-02_analysis.json

Plain files, unencrypted on purpose, in a folder you chose. There is no account for the recorder tier and no server holding the archive — move the folder to another Mac and the meetings come with it. Transcription runs on the machine with Parakeet TDT v3 through Core ML, so no audio is uploaded in any mode; in Pro AI and bring-your-own-key modes the transcript text goes to OpenAI after you explicitly consent, and Pro AI also sends a per-install device id to Narada’s proxy. The recorder tier is free. Pro AI is $6.99/month or $69.99/year with a 3-day trial.

What the audio costs you

A page arguing for retention should do the arithmetic.

  • The playback mix runs about 58 MB per hour. Ten hours of meetings a week is roughly 2.4 GB a month, 28 GB a year.
  • The per-channel microphone and system tracks are kept as cache too, so a recent meeting occupies more than its mix. That cache is deleted after an age you set; the default is 90 days.
  • Older playback files can be recompressed to mono 32 kbps — about 14 MB per hour, fine for re-listening and poor for anything else.
  • “Keep audio for” can be 6 months, 1 year or 2 years instead of forever. Eviction drops the audio and keeps the transcript and summary, so the meeting stays searchable with nothing to play. Favourites are never touched.
  • The on-device AI model, if you use that mode, downloads into the same folder: 3 GB or 7 GB once, depending on your RAM.

If disk is genuinely tight, a cloud notetaker’s free tier storing your meetings on somebody else’s disk is a real advantage of it.

What Narada does not do

  • It cannot recover audio that macOS dropped. Recordings do fail — a device reshaped mid-capture, an overloaded audio daemon, a microphone that quietly stops delivering. Narada compares each finished recording against what was submitted to the file and marks it Incomplete with the reason. That is a warning, not a repair; nothing brings back frames that were never written. An app that never shows you such a warning is not necessarily doing better — it may just not be checking.
  • Speaker labels have no names. The call’s audio is split into Speaker 1, Speaker 2, and the app refuses to guess who they are. Your own lines are attributed to you only because your microphone is a separate file, so an in-person or mic-only recording gets no speaker separation at all.
  • Transcription covers 25 European languages, unevenly. The engine’s published benchmark runs from about 4% word error rate on Italian to about 37% on Greek. Chinese, Japanese, Korean, Arabic, Hindi and Turkish are not supported.
  • macOS 15.6 or newer, Mac only. No iPhone app, no live transcription, no translation, no sharing or team features, no CRM integrations. The Claude Desktop connector is read-only and its search is keyword matching, not semantic.
  • On-device AI is Apple-Silicon-only, produces lighter notes than the cloud route, and pauses while a recording is running.
  • One recording at a time. Two overlapping meetings mean one of them waits.

The part that is not a feature

An audio file of other people talking is a liability as well as an asset, for as long as you keep it. Recording rules vary by country and by state; some places require every participant to agree rather than merely not object. No app makes a recording lawful, and none tells the room on your behalf — Narada shows a reminder on your own screen and a menu-bar indicator, and that is the extent of it.

Say it out loud at the start of the call. And turn on full-disk encryption, since these files are deliberately unencrypted so other tools can read them.

Sources

All read 4 August 2026.

Questions people ask about this

Does Granola save recordings?

No. Granola’s own transcription documentation says “It does not record or save audio or video at any point during the call, so there’s no way to access audio from your meetings.” The same page says it cannot import pre-recorded audio files either, so there is no way to bring your own recording in.

Can I download the audio from Otter?

Yes. Otter’s pricing page lists mp3 among the export formats on every tier including free Basic, so the recording is retrievable. It lives in Otter’s cloud until then, and self-serve retention control is not available below Enterprise.

Where does Narada put the audio file?

In a folder you pick, defaulting to ~/Documents/Narada Meetings, organised by date and meeting name. The meeting folder holds recording_HH-MM-SS.m4a alongside the transcript, the summary and a JSON analysis. Move the folder to another Mac and everything moves with it.

How much disk does an hour of meeting audio use?

About 58 MB for Narada’s playback mix (mono AAC, 128 kbps). Ten hours of meetings a week works out near 2.4 GB a month. Narada can recompress older files to 32 kbps mono, which is about 14 MB an hour, or delete audio past an age you set while keeping the transcript searchable.

Why would I want the audio if I already have a transcript?

Because transcripts are wrong sometimes, and tone is not in them. Checking a disputed quote, re-running a better speech model over an old recording in two years, or handing one file to another tool all need the audio. None of it is possible once the vendor has deleted it.

Try it on your next call

Narada records your Mac's audio and your microphone, transcribes on-device, and hands you the decisions and follow-ups. Free to use, no account, no bot in the meeting.

Download for MacFree · macOS 15.6+ Try the live demo

Keep reading

Alternatives Bot-free AI notetakers — four different mechanisms behind one phrase Four different mechanisms are sold as bot-free, and they have four different privacy stories. Where each one s… Comparison Narada vs Granola — the same bot-free idea, two very different audio paths Both record Mac meetings bot-free. Granola streams audio to Deepgram and AssemblyAI and keeps none; Narada tra… Hub Straight answers about recording meetings Short, sourced answers to the questions people actually ask about recording calls on a Mac: legality, consent,…