Time needed: 90 minutes · 9 steps
What it takes: Software to install
What it costs: A free path exists
Last checked: 2026-09-01

Controlled route for sensitive material

Human rights activist
Journalist or editor
Researcher or analyst

Transcribing testimonies and interviews
Protecting sensitive information
Verifying images, video and claims

These are instructions, not case studies

A recipe tells you what to do. It is not a report of something that already happened somewhere else. Nothing here claims that any organisation did this. It is written for you to do now, and every step tells you how to check it worked.

What you will make

A draft Arabic transcript on an organisation-controlled laptop, linked to the original audio and a human correction log. Use a public or synthetic recording during the workshop.

Choose the lane before copying audio

Public: a published speech can use an approved cloud or local route.
Internal: a staff interview needs an approved organisation account or verified local route.
Identifiable: testimony, protected sources, and recordings that reveal identity or location use the verified local route in this recipe, or are transcribed manually.

Local describes the complete workflow, not just the transcription model. The recording must sit outside OneDrive, iCloud, Google Drive, Dropbox, or any other synced folder. Disable cloud engines, summarisation, and telemetry before protected audio is added.

Workshop route

Install the software and model before any sensitive recording is present.
Create and verify a local, non-synced working folder.
Disconnect the network and prove the selected model still transcribes a test file.
Transcribe the approved recording, then listen back while correcting the draft.

Whisper output is a draft. Published research and the project’s source notes document hallucinated or incorrect transcript text, so any passage used as evidence or a quotation must be checked against the audio.

Five checks before you finish

Data: consent, sensitivity, and the allowed device are recorded.
Input: the correct original recording, speakers, and language were confirmed.
Output: names, places, numbers, dates, and negations were checked by listening.
Human: every quoted passage has a named reviewer and time reference.
Record: the checked transcript, correction log, and original follow the same access rules.

Stop and use another route when

The language is unsupported, the network-off test fails, the folder is synced, or review takes longer than manual transcription. This recipe does not provide a reliable Sorani speech route.

What happens to your information

Never use this for

Do not treat output as testimony, a legal record, a verified quotation or the only copy. Do not use an unencrypted or shared device for material that could harm someone.

What leaves your device

Whose law it lands under · The fully offline option

Software and model download requests leave during setup. A run with the network disconnected cannot send audio over that network. Files, temporary data, backups and the device remain local risks.

Whose law it lands under

Download hosts receive download metadata. No transcription provider receives the interview during a successful disconnected run.

The fully offline option

Install the model, disable analytics and project saving, use a non-synchronised folder, disconnect the network and prove the public test still transcribes.

What it costs

Paying for this from Iraq

No account or card is required. Software and models must be downloaded before offline use.

The free path

If you pay

Vibe and the Whisper models used here have no per-minute fee. Downloads, disk, staff review and a suitable device are still costs.

If you pay

USD 0 for this route. No paid feature is required.

Before you start

A managed laptop with full-disk encryption and enough free space for the model.
Permission to install Vibe and a folder that is not synchronised to cloud storage.
A one-minute public Arabic recording with a known written reference.
A reviewer who understands the speaker and can listen to the original.

Which machines this works on

The Vibe release page currently provides builds for major desktop platforms. Match the asset to the device and read the current release notes. If the processor, signature or organisation policy blocks installation, use an approved machine. Do not bypass a warning you cannot verify.

The steps

Step 1 / 9
Classify the recording. Use a public one-minute known-answer file for setup. Do not copy the real interview to the laptop until the local route passes every check.
Check it worked
The real interview remains in its protected source location and the test file contains no operational or personal information.
Step 2 / 9
Check device encryption, available storage, backups and sync. Create a dedicated working folder outside Desktop and Documents when those folders are synchronised.
Check it worked
The folder path is recorded, sync is off and the recovery key is stored separately from the laptop.

Step 3 / 9
Download Vibe from its official release page and select the asset that matches the operating system and processor. Stop if the release does not support the machine or its signature cannot be verified.
Copy this exactly
```
https://github.com/thewh1teagle/vibe/releases/latest
```
Check it worked
Vibe opens normally and the installed version matches an asset on the official release page.
Step 4 / 9
In Vibe settings, turn off anonymous analytics and automatic project saving. Record the models, projects and temporary-file locations. Do not use summarisation.
Check it worked
The toggles remain off after restarting the app, and the three storage paths are known.
Step 5 / 9
Download a Whisper model using the link documented by Vibe. Set language to Arabic and turn translation to English off. Transcribe the public test file once while online.
Check it worked
The app produces a transcript and the selected model appears in the local models folder.
Step 6 / 9
Disconnect Wi-Fi, mobile hotspot and network cable. Restart Vibe and transcribe the same public file again.
Check it worked
The second transcription completes while the device is visibly disconnected. If it requests a download, the local route is not ready.
Step 7 / 9
Compare the first 100 words with the known reference. Count wrong, missing and invented words. Decide whether the result is usable as a draft for this speaker and recording setup.
Check it worked
The score and go or no-go decision are written in the log. A poor result sends the task to manual transcription or a qualified service.
Step 8 / 9
Copy a working copy of the real interview into the protected folder, keep the master unchanged, and transcribe with the network disconnected. Export plain text and timestamped subtitles.
Check it worked
Both outputs exist, the master audio is unchanged, and no unexpected project copy was created.
Step 9 / 9
Listen from start to finish while correcting the timestamped transcript. Check names, numbers, negation, quotations, silence and speaker changes, then store or delete working files under the project rule.
Check it worked
A reviewer can jump from sampled transcript lines to the matching audio and confirms every sensitive name and quotation.

How to know the whole thing worked

The public file transcribes with the network disconnected, the known-answer score is recorded, and a human verifies the real transcript against original audio with timestamps.

What goes wrong, and what to do about it

The app needs the network after setup. Return to the public test and finish the model download or change route.
The folder syncs to a cloud service. Move it and check backups before using real audio.
The transcript invents speech around silence or noise. Mark the interval and transcribe manually.
The input is Sorani speech. Stop because this Whisper workflow does not support it.

How well it works in your language

Arabic (Modern Standard) · Iraqi Arabic · Kurdish (Sorani)

Arabic (Modern Standard)

Clear read speech is a useful test case, but every transcript still needs listening review.

Iraqi Arabic

Accuracy varies by speaker, dialect, recording, noise and model. No reliable universal score applies to your interview.

Kurdish (Sorani)

Whisper does not list Kurdish in its language table. Do not use this workflow for Sorani speech.

What that rests on

Whisper source lists Arabic but no Kurdish code. Its model card warns about uneven language performance and invented text. Test your device and speakers.

Software named in this recipe

Vibe
Whisper model files
Audacity, optional for a working copy

Whisper on your own computer

Sources

https://github.com/thewh1teagle/vibe/releases/latest · Official Vibe release assets.
https://raw.githubusercontent.com/thewh1teagle/vibe/main/website/public/privacy_policy.md · Vibe privacy policy and analytics description.
https://raw.githubusercontent.com/thewh1teagle/vibe/main/docs/models.md · Vibe model selection and model links.
https://raw.githubusercontent.com/openai/whisper/main/whisper/tokenizer.py · Whisper language table, including Arabic and no Kurdish entry.
https://huggingface.co/openai/whisper-large-v3 · Whisper model card and limitations.
https://arxiv.org/abs/2402.08021 · Primary research on harmful Whisper hallucinations.
https://www.audacityteam.org/download/ · Official Audacity download page.

A CDR original, written for POINT Conference Iraq 7. Checked against vendor pages and package registries in August 2026; the sources are listed on this page.

What we could not verify

This path was not benchmarked on your laptop, microphone, speakers or Iraqi dialect. A disconnected test verifies that route at that moment, not the future security of the device. Sync, backups, exports, repairs or an unlocked laptop can still expose content.

All recipes