Sedjem 0.3.3-preview.20261031

A preview of Sedjem, built automatically on 31 October 2026 from commit a49ad56 on main. It has every change since v0.3.2 (the list). Its tests passed on Windows and Linux, but nobody has tried this build by hand, so for everyday use take the latest release.

• Windows 10 and 11: Sedjem-0.3.3-preview.20261031-setup.exe installs for your user, without admin rights. It isn't signed, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
• Linux (x64): unpack Sedjem-0.3.3-preview.20261031-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

A new preview is built each night when main has changed and its tests have passed. Previews are removed after 7 days, except the newest.

Sedjem 0.3.2

Sedjem 0.3.2 can send a recording to another computer of yours on the same network and transcribe it there, links
each recording's runs from Run again, sorts the list of transcriptions, and makes Update now on Windows wait
for the old Sedjem to quit before it replaces it.

Download

• Windows 10 and 11: Sedjem-0.3.2-setup.exe installs for your user, without admin rights. It isn't
  signed yet, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
  Installing over 0.3.1 keeps your transcriptions, keys and models.
• Linux (x64): unpack Sedjem-0.3.2-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

0.3.1 can update itself: Settings → Updates, then Update now.

Run on another computer

• A laptop without a graphics card can send a recording to another computer of yours that has one, on the same
  network or over Tailscale, and get the transcript back, line by line as it is made.
• On the computer that will do the work, turn on Settings → Other computers → Run jobs for other computers and
  show a pairing code. On the other, type its address and the code once. New transcription and Run again then
  have Run on, and the form asks before the first recording goes to each computer.
• It is off by default. The two computers talk over an encrypted connection, check each other with the one-time
  code and then with the certificate and a token, and only on private addresses. The recording is kept apart from
  the other computer's own library and deleted there once the transcript is back. Everything that travels is in
  the privacy policy, and how it works
  in running jobs on another computer.

The list and Run again

• The list of transcriptions sorts by newest or oldest, title, length, last change or the date the recording was
  made, and remembers the choice.
• Run again with shows which models already ran on the recording and links to their transcripts.

Windows

• Update now no longer fails when the installer reaches Sedjem's files before the old Sedjem has quit: it waits
  for it, up to a minute.
• An installation by hand has a Privacy page after the licence, with what Sedjem sends and the box for the
  anonymous usage counts. Its choice is the same as the switch in Settings → About.
• The installer and Sedjem.exe show their version and publisher in the file's properties.

Other fixes

• A model download that finished while its status was being read no longer showed an error for a moment.
• A copy of Sedjem run from its source code, pulled by hand past the version Update now offered, starts again
  on its own code instead of failing.

Known limits

• Running on another computer is new. Its tests run two copies of Sedjem on one computer; it hasn't been tried
  between two computers on many networks yet. The first time it is turned on, Windows asks whether to let Sedjem
  through its firewall: allow it on private networks only.
• The installer is unsigned, and macOS isn't built.
• On Linux, Update now unpacks the new build next to the old one rather than replacing it.
• Nemotron-3-Diarization hears at most 8 speakers.

Every commit: v0.3.1...v0.3.2.

Sedjem is by Mohammed El-sayed Ahmed and is free software under the AGPL-3.0 (see LICENSE and NOTICE);
commercial licences are available from the author.

The notes behind the numbers are in the project docs.

---

Privacy policy

Sedjem 0.3.1

Sedjem 0.3.1 counts more of how it is used, still anonymously, and no longer loses the counts that waited while the
computer was offline.

Download

• Windows 10 and 11: Sedjem-0.3.1-setup.exe installs for your user, without admin rights. It isn't
  signed yet, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
  Installing over 0.3.0 keeps your transcriptions, keys and models.
• Linux (x64): unpack Sedjem-0.3.1-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

0.3.0 can update itself: Settings → Updates, then Update now.

Anonymous usage counts

• The same three events as in 0.3.0, about six a month for an install used every week, with more in them. Every
  value is still a whole number or one of a few fixed words, and none of them names you, your recordings or your
  transcripts.
• Once a month: how long ago the install was first counted, some settings (the update channel and check, the
  default language, the speaker model, whether the graphics card is used, whether assistants may save), the
  computer in broad ranges (the makers of the processor and the graphics card, the graphics memory, CUDA and
  Vulkan, the power profile, the free disk space), the size of the library in ranges, which models and speaker
  models are downloaded, and which hosted services have a key (whether there is one, never the key).
• Once for each week of use: on how many days Sedjem was used and how often it was started; the runs by hosted
  service, speaker model, length, language and device, with their minutes of audio and of processing; the failures
  by where they stopped, and the cancels; the YouTube captions taken and the courses added; the versions saved by
  editing, the renames, merges, branches, restores and combines; the copies, model downloads and uses of Update
  now; and how many requests assistants made and how many versions they saved, never what they asked.
• The events say which month or week they count and get their time when they are sent, so counts that waited
  while the computer was offline are no longer lost after a day.
• The switch is still in Settings → About, and DO_NOT_TRACK=1 or SEDJEM_NO_USAGE=1 still turns the counts
  off. Everything that is sent is listed in
  usage counts.

Known limits

• The installer is unsigned, and macOS isn't built.
• On Linux, Update now unpacks the new build next to the old one rather than replacing it.
• Nemotron-3-Diarization hears at most 8 speakers.

Sedjem is by Mohammed El-sayed Ahmed and is free software under the AGPL-3.0 (see LICENSE and NOTICE);
commercial licences are available from the author.

The notes behind the numbers are in the project docs.

Sedjem 0.3.0

Sedjem 0.3.0 updates itself, answers assistants such as Claude over MCP, keeps branches of a transcript, and
can tell speakers apart with NVIDIA's Nemotron-3-Diarization. It also sends a few anonymous usage counts, which
can be turned off.

Download

• Windows 10 and 11: Sedjem-0.3.0-setup.exe installs for your user, without admin rights. It isn't
  signed yet, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
  Installing over 0.2.0 keeps your transcriptions, keys and models.
• Linux (x64): unpack Sedjem-0.3.0-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

0.2.0 can't update itself, so this one is installed by hand. From 0.3.0 on, Sedjem tells you when a newer
version is out and updates itself when you ask.

Updates

• On first start Sedjem asks whether it may check GitHub once a day for a newer version. GitHub sees the
  computer's IP address; nothing else is sent. Settings → Updates changes this, and Check now always works.
• Two channels: Stable follows the releases, and Latest follows nightly previews of the newest changes that
  passed the tests on Windows and Linux.
• Update now runs the new installer on Windows, fast-forwards a git checkout and installs changed
  requirements, and unpacks the new Linux build next to the old one. Sedjem then starts again on the same port.

Assistants

• Claude Desktop, Claude Code, Cursor and VS Code can list, search and read your transcriptions and courses over
  MCP, even while the app is closed. Prompts for summaries, minutes, study notes, quizzes and flashcards come with
  it.
• Settings → Assistants gives the snippet to connect each one, chooses what they may see, hides single
  transcriptions, and keeps a log of what they asked. Saving a corrected version from an assistant is off until
  you turn it on.

Branches and a main transcript

• Start a new branch from any version of a transcription, and keep one main transcript for each recording. The
  Branches menu lists them, and the main one is marked.
• Assistants and scripts read the main transcript by default, and a save based on an old version is refused
  instead of overwriting newer work.

Speaker labels

• The speaker model is chosen in Settings → Models: TitaNet-small (the default), TitaNet-large, 3D-Speaker
  CAM++ or NVIDIA's Nemotron-3-Diarization. TitaNet-small is used until the chosen one is downloaded.
• Nemotron-3-Diarization finds who speaks when in one model and hears when two people talk at once. It put the
  fewest words on the wrong person on the test meetings and on two public sets, most clearly in meetings of four
  or more people, at about ten times the size of TitaNet-small and four to six times its time. Sedjem runs it with
  NumPy alone, with no PyTorch. The numbers are in
  speaker labels.

Anonymous usage counts

• Sedjem now sends a few anonymous counts a month to Aptabase, in its EU region: the version, the system in broad
  terms, and how many transcriptions were made with which models. It never sends recordings, transcripts, names,
  file names, links or keys, and no identifier.
• The counts are on by default. The first start, or the first start after this update, says so and has their
  switch, and nothing is sent before that. The switch is also in Settings → About, and DO_NOT_TRACK=1 or
  SEDJEM_NO_USAGE=1 turns the counts off.
• Everything that is sent, and when, is listed in
  usage counts.

Fixes

• Windows: the server keeps a longer queue of connections, so a page no longer fails to load while the server is
  busy.
• Linux on Wayland: the app's window keeps Sedjem's icon in the taskbar.
• A model download that ends while Settings refreshes no longer shows as done before the model is ready.
• History: editing one line no longer marks other lines as changed because their speaker was stored one way in
  one version and another way in the next.
• Running from source on Windows: the PowerShell steps work (.\app.cmd), also with uv.

Known limits

• The installer is unsigned, and macOS isn't built.
• On Linux, Update now unpacks the new build next to the old one rather than replacing it.
• Nemotron-3-Diarization hears at most 8 speakers.

Sedjem is by Mohammed El-sayed Ahmed and is free software under the AGPL-3.0 (see LICENSE and NOTICE);
commercial licences are available from the author.

The notes behind the numbers are in the project docs.

Sedjem 0.2.1-preview.20261029

A preview of Sedjem, built automatically on 29 October 2026 from commit b1f2f40 on main. It has every change since v0.2.0 (the list). Its tests passed on Windows and Linux, but nobody has tried this build by hand, so for everyday use take the latest release.

• Windows 10 and 11: Sedjem-0.2.1-preview.20261029-setup.exe installs for your user, without admin rights. It isn't signed, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
• Linux (x64): unpack Sedjem-0.2.1-preview.20261029-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

A new preview is built each night when main has changed and its tests have passed. Previews are removed after 7 days, except the newest.

Sedjem 0.2.0

Sedjem 0.2.0 transcribes YouTube videos and whole playlists, and plays a recording's video beside its
transcript.

Download

• Windows 10 and 11: Sedjem-0.2.0-setup.exe installs for your user, without admin rights. It isn't
  signed yet, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
  Installing over 0.1.0 keeps your transcriptions, keys and models.
• Linux (x64): unpack Sedjem-0.2.0-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

YouTube videos and playlists

• On New transcription, choose YouTube link and paste the link of a video or a playlist. Sedjem shows
  its title, the number of videos, the total length and which videos have captions made by people, and
  lets you leave some out.
• Each video is downloaded at up to 720p and plays beside its transcript.
• When a video has captions written by people in its own language (Arabic or English), they become the
  transcript right away and no model runs. Otherwise the chosen model transcribes it. YouTube's automatic
  captions are never used.
• A playlist becomes a course: one entry in the sidebar with its progress, and on each lecture's page the
  YouTube link, the previous and next lectures, and a button that opens the course folder.
• Each course is saved in the courses folder of the app's data folder: a Markdown and an SRT file per
  lecture, numbered in playlist order, and course.md with every lecture in one file, ready to paste into
  an assistant for study notes. The files are updated when a transcript is edited.
• YouTube works best with a JavaScript runtime. If none is found, Settings offers to download Deno.

Downloading from YouTube is meant for personal use, under your own responsibility for YouTube's terms.

Video beside the transcript

• Recordings with a picture (MP4, MKV, MOV, WebM, AVI and others) now play their video at the top of the
  side column, and the transcript follows it as it follows the audio. The video can be made larger, put in
  picture-in-picture or shown full screen.
• The video is prepared next to the transcription, so the transcript isn't delayed: copied as it is when
  the browser can play it, otherwise converted to H.264 at up to 720p.

Fixes

• Settings no longer jumps back to the top every second and a half while a model downloads, and an API
  key being typed there is no longer erased (#7).
• The playback speed stays when another recording is opened.
• Running from source: PyAV is pinned below 19, which broke reading recordings for the Whisper models.

Known limits

• A model has to be downloaded before starting, even for videos whose captions will be used.
• YouTube videos offered only in VP9 keep just their sound.
• A playlist is cut at 300 videos.
• The installer is unsigned, there are no automatic updates, and macOS isn't built.

Sedjem is by Mohammed El-sayed Ahmed and is free software under the AGPL-3.0 (see LICENSE and NOTICE);
commercial licences are available from the author.

The notes behind the numbers are in the project docs.

Sedjem 0.1.0

The first release of Sedjem: transcripts of Egyptian Arabic–English meetings and calls, made on your
own computer, with each speaker told apart by voice.

Download

• Windows 10 and 11: Sedjem-0.1.0-setup.exe installs for your user, without admin rights. It isn't
  signed yet, so SmartScreen warns about an unknown publisher: choose More info, then Run anyway.
• Linux (x64): unpack Sedjem-0.1.0-linux-x64.tar.gz and run Sedjem/Sedjem.
• SHA256SUMS.txt has the checksums of both.

On first start, open Settings → Models and download a model: Cohere Transcribe
Arabic (1.6 GB) or the whisper-medium code-switching model (778 MB).

What's in it

• Three local models: Cohere Transcribe Arabic (13.4% word error rate on the Egyptian test set), the
  whisper-medium code-switching fine-tune (18.5%, keeps 90% of English terms in English), and Whisper
  large-v3 with an Egyptian style hint.
• Nine hosted services with your own API key: ElevenLabs Scribe, Google Gemini, Deepgram Nova-3,
  AssemblyAI, Azure AI Speech, Speechmatics, OpenAI, Groq and Mistral Voxtral. The recording is
  uploaded only after you confirm.
• More models to add from Settings, each with its measured results, size and licence.
• Speaker labels by voice (TitaNet-small voiceprints), with the number of people given or estimated.
• The graphics card is used when it helps: Cohere on any GPU through Vulkan (NVIDIA, AMD, Intel),
  Whisper on NVIDIA GPUs after a one-click download of NVIDIA's libraries.
• Playback in step with the text, search, renaming and merging speakers, editing lines, and exports
  to text, SRT, WebVTT, Markdown and JSON.
• Every change to a transcript is kept as a version, with v0 being what the model wrote, so any
  earlier version can be viewed or restored.
• Compare transcripts of the same recording from different models, with the differing words marked,
  keep the better line in each place and combine them into a new version.
• Settings for power mode, processor threads and the graphics card, and speed estimates that learn
  from earlier runs on your computer.

Known limits

• Local transcripts are drafts: on real meetings they differed from ElevenLabs on 44–49% of the words.
• Whisper on CUDA hasn't run on a real NVIDIA card yet; it falls back to the processor if anything
  fails.
• The installer is unsigned, there are no automatic updates, and macOS isn't built.

Sedjem is by Mohammed El-sayed Ahmed and is free software under the AGPL-3.0 (see LICENSE and
NOTICE); commercial licences are available from the author. Versions published before this release
were under the MIT licence.

The notes behind the numbers are in the project docs.