Controlled route for sensitive material

A program you install on your own computer that turns recorded speech into written text.

First thing to tryRecord one minute of invented speech, transcribe it with the internet switched off, then mark every error in names and numbers.

CostThe software and the model weights cost nothing and need no account, so the real cost is a person to install a command-line tool, the model file downloaded before you go offline, and slower transcription on an old computer.

Download it

Do this with it

Transcription

Controlled route

Clean up a transcript without changing what was said

25 minutes · Gives you short sections

Do not paste this into a cloud tool.

Answer in [ANSWER LANGUAGE], even if the material below is in another language.

Below is a machine transcript of a recorded interview, and it stays on my own computer, so work only with what is here and keep the whole answer local. Clean it up for reading.

You may:
- Fix punctuation, capitalisation and the obvious spelling of the names I list below.
- Remove filler sounds and repeated false starts.
- Keep the speaker labels exactly as they are.

You may NOT:
- Complete an unfinished sentence.
- Replace a word you think was misheard.
- Reorder anything.
- Remove a repetition that carries meaning.

Keep every [inaudible] marker and its timestamp where it is.

Give me, in this order:
A. The cleaned transcript.
B. A list of every change you made that was more than punctuation, with the original wording beside it.
C. The lines you could not make sense of, quoted, so a person can listen again.

Names and terms spelled correctly: [LIST THE NAMES AND PLACES]

Transcript:
[PASTE THE TRANSCRIPT]

How to use it

Plain verdict

The preferred transcription lane for sensitive audio when a technical facilitator can install it. Local privacy does not remove transcription errors.

Good for

Interviews, meetings, rough subtitles, time-saving first transcripts, and audio that must not be uploaded to a cloud service.

Avoid for

Assuming every app named Whisper is local, overlapping speakers, noisy recordings, verbatim legal records without correction, or promising reliable Sorani transcription.

Data lane

Local only when the official open source package and model weights run on your computer. Audio and transcripts stay in chosen folders. A website or third-party app using Whisper can upload the recording.

Ten-minute safe trial

Record one minute of invented speech with consent. Transcribe it offline, then mark errors in names, numbers, dialect words, speaker changes, and timing. Delete the test recording if it is no longer needed.

Human check

An authorised listener corrects the transcript against the audio. Protect the transcript as carefully as the recording, and separate speakers manually when needed.

Safer or constraint-based alternatives

For material that can be public, a reviewed cloud transcript may be faster. For printed scans, use Tesseract locally. For Sorani, plan for human transcription unless the exact model has passed a representative test.

What happens to your data

Where this data goes

your own computer

Trains on what you type?

No. Your input is not used to train models.

How long it is kept

When you run the open source software locally, audio and transcripts stay in the folders you choose. Installing a third-party Whisper website or app changes this data lane.

Account required

No account is needed.

Getting it, and using it in Iraq

Access in Iraq

Works in Iraq

The open source package has no account or geographic gate. Download the model before an offline session.

Language quality

No one has published a measurement for this.

Arabic (Modern Standard)
Arabic is included in Whisper's multilingual language list. Kurdish is not listed as a built-in language. Support is not a measured quality result. Test a short sample with a fluent reviewer.
Iraqi Arabic
No current first-party benchmark for Iraqi Arabic was found. Test the exact dialect and task with a fluent reviewer.
Kurdish (Sorani)
No current first-party quality benchmark for Sorani was found. Do not infer quality from general multilingual claims.

Sources

openai/whisperOpenAI
Introducing WhisperOpenAI

Last checked: 2026-09-01