When the audio is not allowed to leave the building¶

Most dictation tools are a convenience. For some work they are a prohibition — the audio contains privileged client discussion, patient information, or a source who agreed to talk on the understanding that the recording stays with you.

For that work the question is not "which tool is most accurate?" but "which tool can I use at all?" Cloud dictation sends the audio to someone else's servers, and that single fact disqualifies it in a lot of professional contexts regardless of how good it is.

YazSes runs the speech recognition on your own CPU. No audio, no transcript, and no metadata is transmitted, because there is nothing to transmit to — there is no account, no API key and no server component.

What it actually does for this kind of work¶

All three are local. The meeting recording is deleted after transcription unless you explicitly ask to keep it ([meeting] retain_audio), and speaker names come from voiceprints you enroll yourself — never from a cloud identity.

Prove it, don't trust it¶

The whole claim on this page rests on "it runs locally", so do not take that on faith from a documentation page. Check it in one command:

docker run --rm --network none \
    -v yazses-models:/models -v "$PWD:/data" yazses interview.m4a

--network none gives the container no route to the internet at all. It still produces a correct transcript. See try it without installing.

On a normal install, the equivalent check is to watch for connections and find none:

ss -tunp 2>/dev/null | grep -i yazses || echo "no network connections"

The one time YazSes needs the network is the first run, to download the speech model. After that it does not, which is why the air-gapped case works at all. The full detail is in the privacy statement, and a machine-readable dependency inventory ships as sbom.cdx.json for reviews that require one.

By profession¶

Legal¶

The recurring blocker is that privileged material cannot be processed by a third party without going through a vendor-assessment process, and often not even then. Local transcription removes the transfer entirely rather than trying to make it acceptable.

Practical notes:

• Add your own vocabulary. Case names, statutes, Latin terms and party names are exactly what a general model gets wrong. yazses vocab add <word> builds a personal dictionary that is fed to the recogniser.
• Use a larger model for anything you will file. small.en is measurably more accurate than the default base.en — see benchmarks — at the cost of speed. For dictation you read back anyway, accuracy is usually the better trade.
• Transcripts are ordinary text files on your disk. They inherit whatever retention and encryption policy you already apply to client files; the tool imposes none of its own.

Clinical and healthcare¶

Same shape, higher stakes. The honest position:

• Local processing means patient audio is not disclosed to a transcription vendor.
• It does not make your setup compliant with anything. Device encryption, access control, audit and retention are all outside this tool's scope.
• Never dictate a dosage or a drug name without reading it back. Numbers and similar-sounding drug names are a known weak point of general speech models, and this one has not been evaluated on clinical vocabulary.

If your institution requires validated clinical documentation software, this is not that. It is a general-purpose dictation tool that happens not to phone home.

Journalism¶

The interest here is usually the source, not a regulator:

• A recording that never leaves your laptop cannot be subpoenaed from a vendor, produced in response to a legal demand to a third party, or exposed in that vendor's breach. It can still be taken from your device — this narrows the surface, it does not remove it.
• yazses transcribe with --diarize tags who said what, which makes a two-hour interview searchable without sending it anywhere.
• It works with no connectivity at all — on a plane, in the field, in a building where you would rather not join the Wi-Fi.

Researchers handling consented data¶

Ethics approvals frequently specify that recordings not be shared with third parties. See the dedicated page on research interviews and ethics approval.

What would make this the wrong choice¶

Stated plainly, because a page like this is worthless if it only argues one way:

• You need a documented compliance attestation. YazSes has none. A vendor with a signed BAA or DPA may be the correct answer even though it processes in the cloud.
• You need certified accuracy for medical or legal transcription. Specialist tools are trained on that vocabulary and are validated for it; this is a general model.
• You need a managed audit trail of who transcribed what. There is deliberately no such logging, because there is no telemetry — which is a feature for privacy and a gap for governance.
• Your language is not well covered. The default models are English; other languages need a multilingual model and accuracy varies. See multilingual dictation.

Getting started¶

1. Install — or try it first with nothing installed.
2. yazses vocab add <term> for the vocabulary your field uses.
3. Set a more accurate model in ~/.config/yazses/config.toml: [stt] model = "small.en".
4. yazses verify to confirm the whole chain works before you rely on it.

Related: private and confidential work · offline meeting notes · comparison with other tools