How Patum works

How Patum works: every step on your Mac

From the sound of a call to a filed note, every step runs on your Mac: recording, transcription, speaker names, the note and its check, and the answers Live context gives you during the call.

  1. 01SoundmacOS
    YouMicrophone
    OthersZoom
    Two tracks, so your words are always yours.
  2. 02Live transcriptNeural Engine
    31:07OthersIf we can't roll it back in minutes, it's not a migration, it's a deploy.
    32:05YouGood. Who wants the load test in Q1?
    Each line about a second after the phrase ends.
  3. 03SpeakersAfter the call
    31:07Mira K.If we can't roll it back in minutes, it's not a migration, it's a deploy.
    34:10Alex L.I'll pull the list of who's still on v1.
    Voices told apart, named from the invite and from voices Patum knows.
  4. 04FactsGPU · Qwen3.5 9B
    DecisionShip the queue migration behind a flagMira K. · 31:07
    ActionList the clients still on v1Alex L. · 34:10
    Open questionWho runs the Q1 load testYou · 32:05
  5. 05The noteGPU · Qwen3.5 9B
    Decisions Load test: no owner was named. Projects/Platform/2026-10-29 Weekly platform sync.md
  6. 06ChecksCode, then the model
    8 facts linked to the second they were said
    Who said what matches the speaker on every quote
    3 left blank, because nobody said them
    1 guess removed: “Alex owns the v1 migration”
    On the note for a day after it's written.
The weekly platform sync from the home page, step by step. A cloud AI writes the note instead only if you turn one on, with your own key.

Where each step runs

Step What happens Where
Recording Your microphone on one track, the call app's sound on another. Patum records only the call apps you allow; a browser only once one of its tabs uses the microphone. Your Mac, through macOS. The permission for other apps' sound doesn't let Patum see your screen.
Live transcript Speech is found and transcribed as you talk. A line appears about a second after the phrase ends. Your Mac's Neural Engine
Speaker names When the call ends, the other voices are told apart, matched to the people in the calendar invite and to voices Patum already knows. Your microphone track is you. Your Mac
The note The transcript is read into facts, the note is written from them, then checked against the transcript. Your Mac's GPU, with a local model
Filing The note becomes a Markdown file under its people, project or client, or in your journal. Patum's own vault on your Mac, or your Obsidian vault (Pro)
Before the call Home shows the day's calls. Each one's briefing is written in the morning from your earlier meetings with these people and, with Pro, the notes you changed recently in your Obsidian vault. Your Mac, with a local model. Home itself runs no model and sends nothing.
During the call Live context searches your notes and code and answers with its source. Your Mac

A cloud AI can take over the note, the briefings and the answers if you turn one on with your own key (below). Without one, nothing in this table uses the internet.

The models

Patum downloads its models once, from Hugging Face, after you install it. Setup shows their size before it starts: about 4 GB on a Mac with 16 GB of memory, and about 10 GB on a Mac with more, which also gets the larger model for notes. A model that isn't on your Mac yet downloads when a feature needs it, during a call too: the feature says what it waits for and goes on once the model is in. Each model is fetched at a fixed version, and every file's size and hash is checked before Patum uses it. Every model is under a license that allows commercial use.

Job Model License Size
Finding speech Silero VAD MIT 2 MB
Transcription (default) NVIDIA Parakeet TDT 0.6B v3: English and 24 other European languages CC BY 4.0 483 MB
Transcription (optional) Whisper large-v3 turbo: 99 languages, among them Armenian, Chinese and Japanese MIT 645 MB
Telling voices apart NVIDIA Nemotron 3 Diarization OpenMDW-1.1 199 MB
Recognizing voices NVIDIA TitaNet-L, WeSpeaker ResNet34-LM CC BY 4.0 103 MB
Notes, briefings Qwen3 4B Instruct, or Qwen3.5 9B on Macs with more than 16 GB Apache 2.0 2.3 GB / 6.0 GB
Live context answers, Ask Qwen3 4B Instruct; Ask can use the 9B in Settings › Models Apache 2.0 the same file
Search Qwen3 Embedding 0.6B, BGE Reranker v2 M3 Apache 2.0 351 MB, 629 MB

On an M2 Pro, Parakeet transcribes an hour of audio in under a minute, about 150 times faster than real time; Whisper runs about 21 times faster than real time.

The language models run with Apple's MLX on the GPU; transcription and speaker models run with Core ML. Models load when a recording or a note needs them and unload when they're idle. Acknowledgements lists every model and library with its authors.

Models
Apple M3 Pro · 36 GB memoryModels use 10.02 GB of 214 GB free
TranscriptionThe default and the fastest, 25 European languages; your vocabulary only corrects names and terms afterwards Parakeet v3 483.1 MB
Speaker labelsNemotron diarization, with your microphone on its own track 302.1 MB
Notes and summariesRecommended for 36 GB of memory Qwen3.5 9B (High Quality) 5.98 GB
Live contextFast enough to answer during a call Qwen3 4B (Balanced) 2.28 GB
AskQuestions across your library, any time Qwen3 4B (Balanced) 2.28 GB
Also available
Whisper large v3 turbo99 languages with numbers as digits in all of them; your vocabulary guides it, so names and terms come out right more often; about a seventh of Parakeet's speed · Downloading 62% Pause
Models run on this Mac’s GPU and never update on their own. Licenses: Apache-2.0, MIT, CC-BY and OpenMDW; commercial use is allowed.
On a Mac with 16 GB of memory, notes use Qwen3 4B as well, and the models take about 4 GB.

How accurate it is

We measured Patum on public test recordings, with the models it uses by default and the same steps it takes after a call.

What we measured Result Recordings
Speech read aloud, in English, German, French, Spanish and Russian About 5 to 6 words in 100 come out different from what was said FLEURS
Who said what in meetings of three or four people, once you name the voices 97 to 98 words in 100 of the transcript go under the right person AMI Meeting Corpus
Who spoke when in TV debates and news, with up to 21 speakers More than 95% of the speaking time goes to the right speaker VoxConverse

Parakeet and Nemotron 3 are NVIDIA's models, and NVIDIA publishes how they score on standard tests. Run inside Patum on the same tests, their error rates are at most one and a half percentage points higher than the ones NVIDIA gives on the model cards for Parakeet and Nemotron 3 (checked November 13, 2026).

These are public test recordings, not your calls. People talking over each other, a microphone far from the speaker and a poor connection all add errors, and nothing here measures the notes. That's why every fact in a note links to the second it was said, and the audio stays on your Mac so you can listen again.

We ran Patum 1.1 as it ships, with Parakeet TDT 0.6B v3 for the words and Nemotron 3 Diarization for the speakers, on an M2 Pro in November 2026. A word counts as wrong when it was changed, left out or added (the word error rate), after both texts go through Whisper's text normalizer, so punctuation, capitals and spellings such as "Mr." and "mister" don't count as errors. The recordings are FLEURS, by Conneau and others at Google; the AMI Meeting Corpus, by the AMI consortium; and VoxConverse, by Chung, Huh, Nagrani, Afouras and Zisserman at the University of Oxford. Each is under CC BY 4.0.

How a note is checked

A small model on a Mac writes plainer sentences than the largest cloud models, and any model can get a name or a date wrong. So Patum doesn't take the model's word for anything. It builds the note from facts it can point to, then checks the result in code.

  1. Reading. The transcript is read into a list of facts: decisions, actions, open questions, updates and risks, each with the time and the speaker of the line it came from. On a 45-minute call this takes the 9B model about a minute.
  2. Writing. The note is written from the transcript with that list beside it. Where the two differ, the model is told the transcript is right.
  3. Cleaning up, in code. A due date the model marked as guessed is replaced with a blank. Every number of two or more digits is looked for in the transcript; a line whose number isn't there is rewritten from what was said, or without the number.
  4. Missed items. Decisions, actions and open questions from the list that no line of the note covers are offered back and added where they fit.
  5. Who said what. Code finds lines that put one person's words in another's mouth. Then every line of the note is read against the transcript lines it most likely came from, and the wrong name, a mixed-up owner or a guessed "he" or "she" is corrected.
  6. Linking. Each part of the note that matches a fact gets a dotted underline. Click it and the popover shows who said it, the words, the second, and Play.

What you see: the note's footer counts its facts ("8 facts, each tied to the transcript"), and for a day after it's written a check card lists what the checks did, including the owners and dates left blank because nobody named them and every line they removed or corrected, before and after. Change a fact's words, in Patum or in Obsidian, and it loses its underline and stops counting as checked; you can put the checked text back.

The checks reduce mistakes; they don't remove them. Only the parts of a note that match a fact are underlined, so a sentence without an underline isn't tied to a moment of the call. The audio stays (as long as you keep it), so a line anyone disputes can be heard again.

Client call — Northwind Oct 29 · 16:00 · 41 min · Northwind
Filed just now · checked against the transcriptClient callUse another template…
7 facts linked to the second they were said
Who said what matches the speaker on every quote
1 left blank, because nobody said itA date for Check the SLA credits with finance
Corrected: “Dana R. · Send the renewal terms · Friday”Now: You · Send the renewal terms · Friday
Promises and next steps
WhoWhatBy when YouFriday Dana R.Check the SLA credits with financeNo date was named
TranscriptOct 29 · 16:00
38:02 Dana R.Good. Then we need the renewal terms in writing.
38:09 YouI'll send them to you by Friday.
38:15 Dana R.And I'll check the SLA credits with our finance team.
The draft gave the renewal terms to Dana. The transcript says you promised them, so the check moved the line and kept Friday, the day you named.

Ten templates

Patum picks a template from the meeting, or you choose one. The template sets the note's sections:

Template Sections
General meeting Summary · Discussion · Decisions · Worth quoting · Action items · Open questions
Stand-up Updates · Team risks · Follow-ups
Project sync Summary · Status · Risks and blockers · Decisions · Discussion · Action items · Open questions
1:1 / Ongoing Conversations Context and outcome · Decisions and commitments · For the next 1:1 · Discussion by topic · Source limits
Retrospective Context · What went well · What didn't go well · Agreed changes · Ideas not adopted
Client call Summary · Client goals · Pains and constraints · Objections and concerns · Promises and next steps · Open questions
Lecture / Study Notes Scope and key lessons · Topic notes · Claims, evidence and caveats · Application · Exercises and next steps · Open questions and source limits
Town Hall / All-Hands Overview · Announcements & News · Progress & Updates · Metrics & Numbers · Highlights & Stories · Q&A · What's Next & Asks
Therapist or doctor session What we talked about · What I understood · What was advised · To try before next time · Questions for next time
Personal note Thoughts · Decisions · Feelings · To do

The note of a therapist or doctor session is written from your side, as the client or patient. Personal note is for voice notes. With Pro, you can write templates of your own.

Speakers and voices

On a call, your microphone and the call's sound are separate tracks, so your lines are always yours. The other voices are told apart after the call (a 30-minute call takes about 16 seconds on an M2 Pro) and matched to the people in the calendar invite. In a room, or in a file you import, there's one track, so Patum finds you by your voice.

To recognize people in later meetings, Patum keeps a numeric voice profile for each person you name or confirm. Voice profiles are biometric data. They're computed and stored only on your Mac and never sent anywhere, including to an AI provider you chose. Settings › Recording › Voice profiles lists them; Forget Voice… removes one person's, Forget all voices… removes every one. If you'd rather Patum not recognize voices across meetings at all, turn off Recognize voices across meetings in Settings › Recording. A team's admins can turn it off for work meetings.

Live context during a call

Every call recording gets This call so far, a running digest of the call written on your Mac: the topic now, what was decided and what's still open. Choose a project in the recording window and Patum also searches that project's notes and code while you talk. Type a question, or turn on Answer Questions Heard in the Call to have the questions people ask answered as they come. The answer names the notes, the code or the moment of the call it used; in our tests it arrives in about two seconds.

Live context runs on the Qwen3 4B model, with the search models above, all on your Mac. It reads your notes and the code repositories you add in Settings › Live context › Code repositories, and with Pro the rest of your Obsidian vault. The questions of a call and their answers are kept in memory while the call lasts and aren't stored after it, except a question you add to your notes. With Cloud answers on (Pro), the question and the passages it cites go to the provider you chose, and the recording window says so ("Answers by Anthropic · search on this Mac").

The network during a call

Once its models are in, Patum makes no network connection while it records, with four exceptions, each one your choice:

  • a cloud AI provider you set for something that runs during a call, such as Cloud answers;
  • your team's relay and our license service for your seat, if you're in a team, and with Enterprise, your company's AI server if your admins named one;
  • a report you send from Help › Report a Problem… or Send Feedback…;
  • fetching a license key you just bought.

If something you use during a call needs a model that isn't on your Mac yet, such as notes search for Live context, Patum shows its size and downloads it from Hugging Face, one model at a time, and Pause download holds it. Once every model it uses is on your Mac, nothing is fetched during a call.

Update checks wait until the recording stops, and an update never installs during a call. The monthly license check happens when Patum starts, never while it records.

Check it yourself

Install a network monitor such as LuLu (free) or Little Snitch, start a recording, and talk for a few minutes. With no provider and no team, and the models Patum uses on your Mac, the monitor shows no connections from Patum until you stop.

What leaves your Mac

The app has no account, no analytics and no tracking. The full list is short, and the privacy policy lists every connection with what it sends.

AudioNever
Voice profilesNever
Transcripts and notesOnly text you send to an AI provider with your own key, or that an AI app you connect reads
ModelsDownloaded once from Hugging Face, after install or when a feature first needs one, then used offline
Screenshots of the call (Pro)Never. With screen moments on, a provider you chose gets the text on them, never an image
Update checksPatum asks patum.nikioth.org for the latest version
Trial and license keyYour license key, a key Patum made for this Mac, and for the trial a one-way hash of its hardware id. Nothing about your meetings
Reports you sendOnly when you press Send in Help › Report a Problem… or Send Feedback…: what you wrote, screenshots you add, diagnostic info and an ID that links the report to your license
Text sent to an AI providerOnly with your own key, for the jobs you choose. Every request is listed in Settings
AI apps you connectPro, if you turn on the MCP server: apps on your Mac such as Claude Code or Cursor can search the meetings in your projects and clients, and may send what they read to their own AI service
TeamOnly in a team: the notes their authors publish, encrypted, through a relay that can't read them on its own; your seat's checks with our license service; to start a free month, the work email you type; and, with Enterprise, the text of work meetings for your company's own AI server to write, if it names one

We train nothing on your meetings. Without a team, nothing from them reaches us. In a team, our relay passes the notes members publish, encrypted with keys made on their Macs, and can't read them on its own. A cloud provider you add works under its own terms.

A cloud AI, if you want one

With Pro you can have a cloud model write your notes and briefings or answer in Live context and Ask, with your own key from OpenAI, Anthropic, Google Gemini, xAI or OpenRouter. You choose the job and the provider. Before the first request for a job, Patum shows what will be sent and where, and asks. Settings › Providers lists every request afterwards. Audio and voice profiles are never sent; the provider gets text, and its own privacy policy applies to it.

Briefings go where your notes go unless you choose otherwise in Settings › Providers: a provider of their own, or your Mac. A provider gets notes from your vault for a briefing only once you allow it to read them.

With Pro, you can also run models with Ollama on your Mac. Patum talks to Ollama only on your Mac itself, so nothing leaves it.

What stays on your Mac, and for how long

  • Recordings are in Movies › Patum, one folder per meeting. Keep the audio as long as you like, or have Patum move it to the Trash after 90, 60, 30 or 3 days, or once the note is written. The note and transcript stay.
  • As long as a meeting's audio is kept, Transcribe Again can transcribe it again with the model chosen in Settings › Models, so meetings already recorded can gain from a better model later.
  • Transcripts, speaker labels, notes and their checks, the search index and your settings are in Patum's container on your Mac.
  • Notes are Markdown files, in Patum's own vault or in your Obsidian vault. If that folder syncs through iCloud Drive or Obsidian Sync, that service keeps copies under its own terms.
  • Require Touch ID to open Patum locks it at launch, when the Mac sleeps or locks, and when its window closes.
  • Keep out of screen sharing, on by default, asks macOS to leave the recording window out of screen shares and screenshots. Try it once with your call app's screen sharing before a call where it matters.
  • Settings › General › Your data has Export library… and Delete all data…. The privacy policy lists what each one removes.

What Patum doesn't do

  • It doesn't decide whether you may record. Many places require everyone's consent. Patum reminds you when a recording starts and gives you a line to paste into the chat ("I'm recording this call on my Mac for notes. Nothing leaves my computer."), but asking is up to you.
  • It doesn't sync your meetings to a cloud, make share links or push notes to a CRM. You share a note as a PDF, Markdown or a drafted email.
  • It runs only on Macs with Apple silicon, macOS 26 and 16 GB of memory or more. There's no iPhone, iPad or Windows version.
  • It doesn't split one recording into several notes.
  • Local models write plainer prose than the largest cloud models. The checks above keep the facts tied to the transcript either way; a cloud provider closes the gap in style, and then the transcript goes to that provider.

Free for Mac, with no account and no minute limits. Pro is free for 14 days, no card.

Needs a Mac with Apple silicon, macOS 26 or later and 16 GB of memory or more. To check yours, choose Apple menu › About This Mac.