v0.4.0: Instant, Self-Healing Transcription

Fresh memos transcribe the moment they appear

v0.3 made new recordings jump the queue; v0.4 removes the wait before the queue even starts. New voice memos now wake the transcription worker the moment they're detected instead of waiting for the next periodic scan, and "Retry" on a failed row goes straight to the front of the queue.

A quieter bug is fixed alongside it: a long transcription could restart from scratch after an app restart, because a stale processing lock from the previous run could send an already-finished job back into the queue loop. The lock is now process-aware — a multi-hour recording survives a restart instead of starting over.

The background helper heals itself

Until now, if the background helper hung or stopped responding, the fix was the oldest one in the book: quit and reopen the app. v0.4 makes that the app's job. The helper tracks per-worker heartbeats and restarts a worker that hangs; the app watches the helper itself and relaunches it if it stops responding. Long jobs also emit a periodic progress heartbeat, so a two-hour transcription that's genuinely working no longer looks frozen.

The CLI imports into your library

transcriber transcribe recording.m4a used to print text to the terminal and walk away. Now it does what the app would do: it imports the file into your My Transcriber folder, creates the transcript note, and streams live progress that the app's Transcribe tab picks up. If no background helper is running it starts one automatically (and shuts it down when idle); if the app is already open it uses that.

Re-running the command on the same file is detected and skipped, so a shell script that loops over a folder is safe to run twice. Want the old behaviour — transcript to stdout, nothing saved? Pass --text-only.

A much smaller download

My Transcriber no longer bundles the photo-captioning engines it never uses — the DMG is dramatically smaller and installs faster. Nothing else changes: same all-local Whisper transcription, same data folder.

The microphone works again

A missing hardened-runtime entitlement was silently blocking microphone access for in-app recording — no permission prompt, no entry in Privacy & Security, nothing to toggle. The failure mode was maximally confusing because macOS never said no; it just never asked. It's fixed: recording now requests and gets mic access properly.

Smaller things

• Re-transcribe in the right language. A memo that came out in the wrong language can be re-transcribed with the language forced, from the row's context menu.
• A transcript quality guard. If Whisper goes off the rails (e.g. a hallucinated repetition loop, now detected), the transcript is flagged with a quality note instead of being trimmed, and the engine restarts clean for the next job. Audio is also enhanced before transcription for better results on quiet or noisy recordings.
• A resource history chart. The Processing tab now graphs the background helper's memory and CPU over time, so "what is it doing right now" is answerable at a glance.
• First launch can no longer race the model download. Transcription waits until the Whisper model is fully downloaded — previously a first run could fail confusingly against a partial model — and onboarding shows live download progress.
• Processing tab polish. The "now processing" text no longer spills over the rows below it, and stale processing entries clear correctly when the app reconnects to the helper.

What's next

The theme this cycle was transcription you don't have to babysit — record, walk away, and the transcript is there. The next round builds on that. If you hit anything, bug reports from Settings → Bug report (or via Claude Code through MCP) go straight into the loop.