Guide

Transcribing an interview offline, without sending the audio to the cloud

Page updated

The essentials

To transcribe without exposing a source, run the speech-recognition model on your own device: Whisper, an open-source model, makes that possible. Download it before you leave, record cleanly with the person’s consent, then proofread the transcript, especially proper names.

Why transcribe on the device

Most transcription services upload the audio to their servers. For an ordinary meeting that is acceptable; for a source taking a risk by talking to you, their voice and words end up with a third party. Local transcription removes that transfer, and works without a network.

Step by step

  1. Get consent

    Say you are recording, what the recording is for, and offer anonymity when needed.

  2. Record carefully

    A quiet place, the phone close to the person, not on a vibrating table. Good audio beats a big model.

  3. Download the model before you leave

    The model downloads once, on Wi-Fi; after that it runs without a network.

  4. Transcribe and proofread

    Models mostly get proper names, numbers and mixed languages wrong. Listen again to the passages you will quote.

  5. Anonymize before sharing

    Replace names and identifying details in the text before it leaves your device.

Which model to choose

  • Tiny (about 45 MB): the fastest, recommended on phones.
  • Base (about 90 MB): a good balance, recommended on computers.
  • Small (about 250 MB): more accurate, slower, for a recent computer.

With nTilia

From Pro, nTilia transcribes a note’s audio with Whisper on the device: the audio is not uploaded. The transcript lands in the note with [mm:ss] timestamps; the Outline tab lets you jump to them. On phones, you need a WebGPU-capable browser.

Frequently asked questions

Can I transcribe an interview without an internet connection?

Yes, with a model like Whisper that runs on the device. You just need to have downloaded it beforehand, while you had a network.

Is local transcription as accurate as the cloud?

Small models, suited to phones, make more mistakes than the large server models. Good audio and proofreading make up for most of it.

Is Whisper free?

Whisper is an open-source model released by OpenAI. In nTilia, local transcription is included in the Pro plan.

Sources

Free · No credit card · Works offline