Arm Voice Ledger

• 

  Mobile proof: the same local evidence-to-decision flow.

• 

  Desktop proof: select timestamped evidence, save an editable row, replay source.

• 

  One local voice note becomes a replayable, human-owned decision.

Inspiration

Voice notes are fast to create and easy to lose. A summary may tell you what was said, but it often separates a decision from the exact moment that supports it. I wanted a smaller, more trustworthy loop: preserve only the useful moment, let a human name what it means, and make it replayable.

What it does

Arm Voice Ledger turns one short local WAV into timestamped words. A person selects the exact range that matters and saves it as an editable Decision, Next action, or Question. Every saved row can replay its original source moment in the current browser session, and the ledger can be explicitly exported as local JSON.

It is deliberately not a generic meeting summarizer. The product is about preserving the link between a human decision and the evidence that earned it.

Why Arm

The workflow runs MLX Whisper locally on Apple-silicon Arm64. After an intentional first cache warm, the normal runner forces the model cache offline. There is no application-server audio path: the local server binds to 127.0.0.1, uses a temporary request file, and removes it after transcription.

That makes the Arm contribution concrete: local inference produces word-level timing that becomes an interaction a person can use, not a metadata badge hidden in a log.

How we built it

Python powers the local server, MLX Whisper transcription, timestamp extraction, and smoke checks. The browser surface handles evidence selection, editable ledger rows, source replay, and a local JSON export. The public MIT repository includes setup instructions, a public-safe synthetic fixture generator, and an offline replay command.

The checked proof ran on Apple-silicon M2 Pro and M4 machines. The current bounded replay produced 16 timestamped words from the cached offline model, preserved the local source replay, and reported that request audio was not persisted.

What a judge can reproduce

1. Clone the public MIT source.
2. Generate the included public-safe fixture.
3. Warm the declared model once on Apple silicon.
4. Run the documented offline smoke command and inspect timestamped evidence, an editable row, source replay, and local export.

The public demo shows the same source -> timestamp -> decision -> replay flow in under one minute.

What we learned

On-device inference becomes more valuable when a person can return to the source, correct the interpretation, and carry forward only what matters. The human owns the decision; the model supplies bounded, replayable evidence.

Honest limits

This proof accepts short WAV files up to 8 MB and uses a short English synthetic fixture. It does not claim noisy-audio accuracy, multilingual quality, long-recording support, battery performance, or an iOS/Android release. Reproduction is scoped to the documented Apple-silicon path.

Built With

• machine
• python