Skip to content
Cameron Mills

All projects

Turning speech into filed work, with the source of every transcript on record

The Life Tracker voice layer

Part of Life Tracker

Voice notes from the phone, a call recorder that transcribes on the Mac, and Wispr Flow as a voice layer, all feeding one system that records what it did with each note and where every transcript came from.

Code on GitHub

How it works

Two paths from speech to filed work: voice notes through an iCloud drop folder, and calls through the recorder. Select a layer to open it.

Problem

I talk to my Life Tracker as often as I type. That means voice notes from the phone, calls to transcribe and spoken answers in practice sessions. Each of these had to run unattended, record where it came from, and keep a note safe when something stalls.

Approach

Keep the unattended path local and mechanical, and add Wispr Flow where it helps, with the local engines as a fallback. No program can start Flow's dictation or read it back, and its meeting data is only reachable from a session I am in, so the design works within those limits.

Architecture

Voice notes. A watcher on an iCloud folder picks up a new note within seconds, with the scheduler as a backstop. A step with no AI in it turns each note into a candidate, and then a small agent run sorts each one into a reminder, a question, a task, a piece of work or a plain note. An audio note is only accepted once its file has been checked and decodes cleanly.

Call recorder. A small Mac program waits for a call app to take the microphone, then records both sides separately, transcribes each on the Mac and merges them, removing echo. Nothing the recorder captures leaves the Mac unless I explicitly send it.

Wispr Flow. When Flow is running, the practice chat's microphone button hands over to Flow, which types into the answer box, and otherwise a local speech model takes over. For calls, the recorder matches a Flow meeting to a recording by time. Flow's transcript is used, the local one is kept beside it, and every transcript records where it came from. Only Flow's meeting records are read, never its dictation history, and a test checks that.

What I built

Notes once sat for hours on files iCloud had moved off the Mac, and two copies of the sorter could run at once. Such notes are now fetched with a time limit, and a proper lock stops duplicates. A note that keeps failing backs off and, after repeated attempts, stops with a single notification. A receipt in iCloud lists what each recent note became. The other direction is a spoken good-morning, which since 14 October opens with the weather, the important email since the evening before and any LinkedIn messages, naming only people and topics.

Results

Measured live on 11 September through the iCloud folder watcher, typed and dictated text notes went from the phone to a filed result in under a minute and a half. The recorder has its own test suite, including cases for Flow, and a separate test covers the microphone with Flow running, stopped or unknown.

What didn't work

An unanswered microphone permission prompt once froze the recorder, so start-up now has watchdogs. No real Flow meeting result has been seen yet, so that part is only tested against assumed formats. A meeting Flow records is transcribed on Wispr's servers, so the local-only rule does not cover it. Audio notes are built but switched off, and the latest voice changes have not yet run live.

Stack

  • Python
  • Swift
  • Core Audio
  • Apple Speech
  • Whisper
  • Wispr Flow

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit c8e90f7 on 1 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.