Skip to content

Private dictation for confidential work

Short answer: if your audio must not leave the machine — because of a patient, a client, a source, an NDA, or a security policy — cloud dictation is not an option no matter how accurate it is. YazSes transcribes on your own CPU, with no account and no network call, so there is nothing to leak.

The problem with cloud dictation

Every mainstream dictation tool — Google Voice Typing, Wispr Flow, most phone keyboards, and the "AI notetaker" category — works by streaming your microphone to someone else's server. For a lot of work that is simply disqualifying:

  • Clinical — dictating patient notes to a third-party processor creates a disclosure you have to justify and a processor you have to have an agreement with.
  • Legal — privileged material passing through an external service is a risk many firms will not accept.
  • Journalism & research — source protection is incompatible with an upload you do not control.
  • Corporate & government — machines that are network-restricted or air-gapped cannot reach a cloud API at all, so cloud tools do not merely violate policy, they do not function.

What YazSes actually does

Property YazSes
Where audio is transcribed On your CPU, locally
Network required after install None
Account / API key / subscription None
Telemetry None
Always listening No — hold-to-talk; the mic is only live while you hold the key
Where transcripts go Typed into the focused app; not stored unless you opt in

The model file is downloaded once during setup. After that you can unplug the network permanently and dictation continues to work. This is verifiable — the project is Apache-2.0 licensed and the source is on GitHub.

Push-to-talk, not always-on

This is a meaningful design difference from "AI assistant" style tools. There is no wake word listening in the background by default and no continuous capture. The microphone is only recording during the window in which you are physically holding the hotkey down. When you release it, capture stops.

What is stored, if anything

By default: nothing persistent. The transcript is typed and forgotten.

There is an opt-in on-device learning corpus that stores your dictations so that yazses tune can propose accuracy improvements from your own corrections. It is off unless you turn it on, and when on:

  • text and audio are encrypted at rest with a machine-bound AES-256-GCM key;
  • it never leaves the machine;
  • you configure regex patterns that are scrubbed before anything is written;
  • retention and size caps evict old data automatically.
yazses corpus status    # what has been captured, if anything
yazses corpus destroy   # delete it all

Full details in the privacy statement.

Air-gapped and restricted machines

Because there is no licence server or activation call, YazSes runs on a machine that has never touched the internet, provided you carry in the package and the model file. Both the snap and the PyPI wheel can be transferred offline, and the model can be pre-placed in the model cache directory.

Confidential meetings

The same guarantees extend to meeting capture and to transcribing existing recordings — both run through the same local pipeline. If your meetings cannot be sent to Otter.ai or Fireflies, see:

Honest limits

  • Accuracy is not magic. A local CPU model is very good, but for professional medical or legal dictation with specialist vocabulary, a mature commercial product like Dragon still has an edge. You can close part of that gap with a personal vocabulary and the opt-in learning loop, but be realistic.
  • "Offline" is about the audio, not about the operating system. YazSes cannot stop other software on your machine from doing whatever it does.
  • This page is not legal advice. Whether a given workflow satisfies HIPAA, GDPR or your institution's policy is a determination for you and your compliance function. What YazSes provides is the technical property those frameworks usually care about: the audio is never transmitted.