gov2ai MENAPOINT Conference Iraq 718–20 September 2026 · Erbil, Iraq

Whisper on your own computer

Turns recorded speech into text without the recording ever leaving your machine.

ToolsOpenAI (model weights, MIT licensed)github.comOpen sourceMITCan run on your own serverWorks offline

Safe for sensitive material

Whisper turns recorded speech into text. You download it once and run it on your own laptop, so an interview recording never leaves the room it was made in.

It comes in several model sizes. The small ones are quick and make more mistakes. The large ones are noticeably more accurate and, on a machine without a dedicated graphics card, can take longer to transcribe a recording than the recording itself lasts. Start a batch of interviews before lunch, not before a deadline.

Expect to correct names, places, and anything two people said at once.

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

Nothing is sent anywhere. Files stay where you put them.

Account required

No account is needed.

Getting it, and using it in Iraq

Access in Iraq

Works in Iraq

Downloads once, then works with no internet connection.

Language quality

  • Arabic (Modern Standard)No one has measured this

    No one has published a measurement for this.

    Arabic is one of Whisper's 99 published supported languages (https://github.com/openai/whisper/blob/main/whisper/tokenizer.py), and OpenAI publishes a per-language WER/CER chart for the large models in the Whisper paper and README, but no specific Modern Standard Arabic figure has been read off that chart or independently benchmarked for this portal. Untested until someone does.

  • Iraqi ArabicNo one has measured this

    No one has published a measurement for this.

    No dialect-specific benchmark exists. The Common Voice and FLEURS test sets used in Whisper's published evaluations are read-aloud Modern Standard Arabic, not spoken Iraqi colloquial Arabic, so a strong MSA score would not carry over even if one were verified. Untested.

  • Kurdish (Sorani)Poor

    Kurdish, in any variant including Sorani, is absent from Whisper's own list of 99 supported languages: https://github.com/openai/whisper/blob/main/whisper/tokenizer.py (LANGUAGES dict, checked 2026-08-30). A language the model was never trained to recognise should be expected to transcribe very poorly, if at all.

Sources

Last checked: 2026-08-30