Meetings don’t speak one language.

Sedjem turns Egyptian Arabic–English meetings and calls into transcripts on your own computer. The English words stay in English, and each speaker is told apart by voice.

Download for Windows Linux, and the source

0:00Mona
صباح الخير يا جماعة، خلونا نبدأ ال sprint planning بسرعة عشان عندنا demo الساعة اتنين
0:05Karim
تمام، أنا بس عايز أفكركم إن ال release بتاعة الأسبوع اللي فات لسه فيها bug في ال login
0:10Mona
The login bug is on the board already, صح يا كريم؟
0:14Karim
أيوه، عملتله ticket وحطيته high priority، المشكلة في ال session timeout
0:19Omar
أنا شفت ال logs امبارح، ال token بيخلص قبل ما ال refresh يحصل، فال user بيطلع برا
0:25Mona
طب مين هياخده الـ sprint دي؟

sprint-planning.srt
1
00:00:00,400 --> 00:00:05,000
Mona: صباح الخير يا جماعة، خلونا نبدأ ال sprint planning بسرعة عشان عندنا demo الساعة اتنين

2
00:00:05,300 --> 00:00:10,500
Karim: تمام، أنا بس عايز أفكركم إن ال release بتاعة الأسبوع اللي فات لسه فيها bug في ال login

3
00:00:10,800 --> 00:00:14,200
Mona: The login bug is on the board already, صح يا كريم؟

A made-up sprint planning as Sedjem writes it, and its first lines exported as subtitles.

English words stay English

Egyptian teams talk in Arabic sentences full of English words. Many speech models write those words in Arabic letters; the ones Sedjem runs keep them in Latin script.

As some models write it: صباح الخير يا جماعة، خلونا نبدأ السبرينت بلانينج بسرعة عشان عندنا ديمو الساعة اتنين
As Sedjem writes it: صباح الخير يا جماعة، خلونا نبدأ ال sprint planning بسرعة عشان عندنا demo الساعة اتنين

One line, written both ways.

Checked on a public test set

On 40 public Egyptian Arabic–English clips, several models kept only 8–25% of the English terms in Latin script. The ones Sedjem runs kept 77–90%.

Real meetings are harder than any test set, so treat a local transcript as a draft. For minutes that people will rely on, a hosted service is still more accurate.

The method and every number

Measured on Perle, a public set of 40 Egyptian Arabic–English clips, on the project’s test laptop: an Intel Core i5-1245U with no NVIDIA graphics card.
Model	WERlower is better	English kepthigher is better	1 h takeson the test laptop
Cohere Transcribe Arabic	13.4%	77%	about 20 min
whisper-medium code-switching default	18.5%	90%	30–45 min
Whisper large-v3 with an Egyptian style hint	20.8%	88%	about 2 h

Every voice gets its own name

A voiceprint model tells the speakers apart and labels their lines; you name each one once.

In the project’s tests on meetings of two or three people, 8–9% of the words went to the wrong person, against 6% for ElevenLabs’ own labels.

How the speaker labels work

The app’s Speakers panel for the made-up sprint planning: Mona 38%, Karim 22%, Omar 22% and Nour 17% of the talking, each name with a Merge into button — The app’s Speakers panel, with each voice’s share of the talking.

Your recording stays on your computer

The models run on your own processor or graphics card. There is no account, and a recording is uploaded only if you choose a hosted service and allow it, each time.

Sedjem also sends a few anonymous usage counts, such as its version, your country and how many transcriptions it made, so the project knows where it’s used. Never your recordings, transcripts or file names. You can switch them off in Settings. The website itself counts page views and download clicks anonymously, without cookies, and honours Do Not Track.

Privacy, in full

The app’s Settings, Models on this computer: whisper-medium code-switching, taking 778 MB, and Cohere Transcribe Arabic, taking 1.6 GB, both downloaded, and Whisper large-v3 still to download — The models on this computer, as the app’s Settings show them.

Click a line to hear it

The transcript follows the recording as it plays. Search it, rename the speakers and correct the text while you listen.

The transcript in the app: lines of Arabic with English terms, each with its time and speaker, and the line being heard marked — The transcript, with the line being heard marked.

Every fix is a new version

Edits are shown as a diff before they’re saved, and the model’s own output stays as version 0.

Review your changes in the app: Omar’s line as it was, in red, and as it is now, in green, with eight story points marked — One edited line, reviewed before it is saved.

Two models, the better line from each

Run a second model on the same recording. Sedjem lines the transcripts up, marks the words that differ, and keeps the line you choose.

The app, step by step

Compare transcripts in the app: Omar’s line as Cohere and whisper-medium wrote it, the differing word marked, and whisper-medium’s line kept — One line from two models, with the word that differs marked.

Runs on an ordinary laptop

No graphics card is needed: on the project’s test laptop, Cohere Transcribe Arabic writes up an hour of audio in about 20 minutes.

On your computer

Platforms: Windows 10 and 11, Linux on x64
Graphics card: Optional, it speeds things up
Download: About 80 MB, then a model of 778 MB or 1.6 GB

With your recordings

Languages: Egyptian Arabic and English, mixed
Opens: Audio and video files, YouTube links and playlists
Exports: Plain text, SRT, WebVTT, Markdown and JSON
Assistants: Claude, Cursor and VS Code, over MCP

Licence: AGPL-3.0, free and open source

Get Sedjem

Free and open source under the AGPL-3.0. Version 0.3.1.

Windows

Windows 10 and 11. The installer is about 80 MB and installs for your user, without admin rights. It isn’t signed yet, so Windows SmartScreen warns about an unknown publisher: choose More info, then Run anyway.

Download for Windows

Linux

Linux on x64. Unpack the archive and run Sedjem/Sedjem.

Download for Linux

From source

With Python 3.12, Git and uv:

git clone https://github.com/MohammedEl-sayedAhmed/sedjem.git
cd sedjem
uv venv .venv --python 3.12
uv pip install --python .venv -r requirements.txt
./app.sh

Windows and other setups

First start

In Settings → Models, download whisper-medium (778 MB) or Cohere Transcribe Arabic (1.6 GB). The speaker model, TitaNet-small (40 MB), comes with the first one. A graphics card is optional.