←  Back to Research and Thoughts
AllSaid

How it works

Record all day. Hold almost nothing in memory.

An app that runs through your whole working day has to cost close to nothing while it waits. So AllSaid is split in two: a small recorder that stays, and a worker that does the heavy lifting between calls and then quits.

The path of a call

Your voice is recorded from the microphone and the far end straight from the call app, as two separate tracks. That separation is why “Me” in a transcript is a fact rather than a guess: only the far-end track is ever split into Speaker 1, Speaker 2 and so on.

The worker loads a speech model, clears the queue, and exits. Exiting is the memory management — between batches it doesn’t exist at all.

Measured

Numbers, not adjectives.

PartMemoryWhen it exists
Menu-bar recorder~45 MBAll day. Loads no speech model of any kind.
Worker, English run~165 MB peakOnly while transcribing, then it exits.
Summariser (default)0 MB hereApple’s on-device model runs in the system’s own process.

Speech models

English out of the box. Spanish when it’s spoken.

LanguageModelWhy
EnglishParakeet Unified 0.6B, int81.82% word error rate at 123× realtime, with punctuation
SpanishNemotron Multilingual 0.6BCovers the language and punctuates

With Detect for each call, the first 30 seconds of a session are listened to before any model is chosen. Detection runs on the audio rather than the text — Spanish run through an English model doesn’t come out as Spanish, it comes out as confident English nonsense.

Settings that matter

The decisions are yours, and the defaults are cautious.

Pause
One click in the menu bar stops all recording. The icon always shows whether it is recording, paused, or capturing only one side.
Never record these apps
Anything currently using the microphone is offered for one-click blocking.
Start at login
Off until you turn it on. An app that records your day should be a choice you make.
Audio retention
0 to 7 days. Ships at 0.
Summary language
Whatever was spoken, always English, or always Spanish. Transcripts are never translated.
Summariser
On-device by default; local runtimes such as Ollama, or cloud providers if you opt in.
←  Back to Research and Thoughts