Voice notes and recordings

Typing and handwriting both need your hands and your attention. Talking doesn’t. Pellet records straight into your vault, plays audio anywhere you can embed a file, and can turn a recording into text you can search.

Recording

On desktop, press the microphone button in the left toolbar, next to New. On a phone or tablet, choose New Recording from the new-document menu.

A small pill appears while you’re recording, showing how long you’ve been going. From it you can:

• Pause and pick up again
• Stop and save
• Discard — which sends the part-recording to the trash rather than deleting it outright

The pill stays with you. Wander off to another note, open a board, collapse the sidebar: the recording keeps going. When you stop, the recording is filed in your Attachments folder and a link to it drops into the note you were in when you started — even if that’s not the note you’re looking at now.

Start a recording with no note open and it still saves. Pellet just tells you the filename instead.

On a phone, recording carries on in the background and from the lock screen.

What you get

An .m4a file — mono, about 21 MB an hour. It’s an ordinary audio file: double-click it in Finder, open it on your phone, email it to someone. Nothing about it needs Pellet.

That’s a deliberate trade. There are formats that would halve the size, but they don’t play everywhere yet, and a recording you can’t open in five years isn’t worth the saved megabytes.

Playing

Audio embedded in a note plays right there — in the editor and in reading view — with a waveform you can scrub, elapsed and total time, and a speed control that cycles through 1×, 1.25×, 1.5×, 2× and 0.75×. It works the same on a canvas, on a whiteboard, and when you open the file on its own.

Drop any audio file into your vault and it behaves the same way: .mp3, .wav, .m4a, .aac, .ogg, .opus, .flac, .aiff.

Playback starts immediately and you can drag to any point straight away, even in an hour-long recording on a phone — Pellet fetches the part you asked for rather than the whole file.

Turning a recording into text

A recording is a black box: you can’t search it, nothing links to it, and it doesn’t appear in your graph. Transcribing fixes all three at once.

Press Transcribe on the player. Pellet writes a new note under Transcripts, with a link back to the recording. From that moment it’s an ordinary note — it turns up in search, in the quick switcher, in backlinks, and on the graph.

Whether the transcript carries timecodes down the side, matching the player’s, depends on what the service sends back. Some return the text broken into timed segments and Pellet lays those out as timecoded lines; others return the words and nothing else, and you get continuous prose. Gemini returns no timings, and neither does OpenAI’s gpt-4o-transcribe. Don’t count on timecodes until you’ve seen them for the model you picked.

This uses a service you choose in Settings → AI & MCP, with your own API key unless you run it yourself (the tab is shortened to AI on a phone). Audio has its own Audio transcription card there — it’s separate from Paper transcription, so setting up handwriting doesn’t set up audio, and the two can use different services.

You have to pick a model. There is no default, deliberately: a baked-in model name goes stale without anyone noticing. Until you choose one, transcribing stops and tells you no model is set. Choose OpenAI or Google (Gemini) as the provider, then a model. Nothing is sent anywhere until you press the button.

On a phone, choose Google (Gemini) — the way OpenAI wants the file uploaded isn’t something a phone can do.

Keeping it on your own machine

There is a third option: Local Whisper server. Run whisper.cpp’s server on a computer you own, put its address in Settings (blank means http://127.0.0.1:8080), and recordings are transcribed there instead of by anyone else. No API key, and no model to choose here — the server is started with the model you want.

It works from a phone too, as long as the phone can reach that computer: use the machine’s address on your network rather than 127.0.0.1, which to a phone means the phone.

The address is remembered per device, for that reason — the same server is 127.0.0.1 to itself and a network address to everything else.

Action items become tasks

This is the part worth the trouble.

Open a transcript and Pellet tells you how many lines in it sound like something somebody agreed to do. Press Find Action Items and you get a list: what it thinks the task is, and the line it came from, so you can judge it without going back to the audio.

Everything is ticked to start with. Untick whatever isn’t a real task and press Create.

Nothing is created until you say so. The list is found by looking for the shapes commitments take — “I’ll send…”, “can you…”, “action item…” — and it errs on the side of offering too much, because a missed one is gone and a wrong one costs a glance. It’ll suggest things that aren’t tasks. That’s the design, not a fault.

Give any of them a date while you’re there: the calendar button on a row opens a date picker, and the date appears next to the task. Click the button again to change it. If you wrote “(scheduled)” next to something on a handwritten page, Pellet takes the hint and highlights that row’s button. It leaves your words alone — what was transcribed is what you get, and you can edit the title yourself if you want it shorter.

Once you’ve created the tasks the bar goes away, so you’re not offered the same ones twice. You can also close it yourself with the × at its right-hand end.

This works on any note, not just a transcript: the Find Action Items button in a note’s toolbar — the circled checkmark — reads the note the same way. Anything you’ve written as an open - [ ] is offered whatever it says — a checkbox is already you saying it needs doing. Ticked ones are left alone.

The line in the transcript turns into a link to the task it became, so the page you were reading and the task list stay joined up — the same thing that happens when you convert a checklist item.

Run it again later and anything you’ve already turned into a task is left out, so you never get the same task twice. Lines you’ve added since are offered as usual.

Your words are never removed. A line is only replaced by its link when the link carries the whole line — in which case the task is named exactly what the line said. If the line says more than the task’s name can (a second sentence, an awkward character, or simply too much text), the line stays exactly as you wrote it and the link is added on an indented line underneath.

Each task you keep quotes what was actually said and links back to the recording. Where the transcript has timecodes, the link lands at the moment it was said — so six weeks later, when the task makes no sense, one click plays you the thirty seconds that explain it. Without timecodes, the link takes you to the recording and you find the spot yourself.

Exporting

Export a single note with Export… in its ⋯ menu, or read a folder as one document and export the lot. Both offer the same five formats, and they treat recordings differently:

• Web page (.html) carries a working player, for recordings up to about 6 MB. Longer ones are named instead of embedded — otherwise a handful of recordings would make a file too big to email. This is the only export you can actually play.
• Markdown keeps the recording embedded exactly as you wrote it, so it still plays wherever the vault is.
• PDF names the recording but cannot play it. The PDF is that same web page printed, and print freezes a player into a picture of one.
• RTF and plain text name the recording too, since neither can play anything.

A word on space

An hour a day is roughly 7.5 GB a year. If you sync your vault between devices, recordings sync too — worth knowing before you record every meeting for a year.

Permissions

The first time you record, your operating system asks whether Pellet may use the microphone. Say no and recording won’t work; you can change your mind in your system settings.