Transcription that never leaves the machine

For audio you are not allowed to upload. Runs on local hardware, offline, with no per-minute billing — and I have the measured receipt.

Cloud transcription sends the recording to someone else's servers. For litigators, clinicians and government teams handling privileged material, that is a compliance problem, not a convenience. This runs with the Wi-Fi off. Nothing uploads, nothing is billed per minute, nothing is retained by a vendor — because there is no vendor in the loop.

The measurement (mine, not a brochure)

# 10.17 min of speech, 16kHz mono, on a Mac, CPU only, network off
$ python3 transcribe_local.py proof.wav --model tiny.en
# language=en duration=610.3s model=tiny.en wall=37.3s (local, no network)

16.4x realtime

610 seconds of audio in 37 seconds of wall clock, on CPU. One hour of audio costs about 3.7 minutes of machine time.

$0 per minute

No API, no per-minute billing, no upload, no seat minimum. Open-source models on hardware that is already on the desk.

Accuracy, honestly

Names, an address ("1142 West Palm Avenue"), times and an invoice complaint all came through intact. Punctuation needed a pass.

The caveat

That run used the fast model. Noisy rooms want a larger one, which trades speed for accuracy. A real engagement runs your worst-case recording before anything is promised.

Who this is for

BuyerWhy it lands
Litigation firmsPrivileged recordings, depositions and client calls cannot be shipped to a US cloud on a whim. Bar associations are publishing on exactly this risk.
Clinics and practicesPatient audio is regulated. Local transcription keeps the recording inside the practice.
Government and agenciesResident and caseworker audio stays on agency hardware, with an audit trail nobody else holds.

How an engagement works

1. Send one hard recording

Your worst-case file — accent, crosstalk, phone quality. I run it locally and send back the transcript.

2. You judge the output

If the accuracy is not good enough for your work, we stop there. No fee, no obligation.

3. If it holds up

A flat monthly arrangement for the volume you actually handle. No per-minute meter, no usage spike surprises.

Send me one recording to test

Related

I also track what every AI model costs per million tokens, and what moved today — AI Price Watch → — and run a free 30-minute session that cuts a business's AI bill, done live: the AI Bill Clinic →

Everything on this page was measured on the machine this page was built on. The script, the model list and the sample clip are not client material — nothing confidential is here, by design. Re-runnable in full.