Şerḥ — Intelligent Ottoman Manuscript Transcription
Şerḥ is a browser-based workspace, with an optional desktop app, for annotating, transliterating, and translating Ottoman manuscripts and archival documents. Developed out of my own research workflow, it is designed to bring manuscript images, word processing, and dictionary tools into a single environment, with optional AI assistance for reading and interpretation.
Rather than moving continuously between an image viewer, a word processor, online and printed dictionaries, and separate AI interfaces, Şerḥ keeps the manuscript page and the text being produced from it together. The aim is not to automate manuscript reading, but to make the scholarly workflow around it more coherent: defining the page, working line by line, checking uncertain readings, consulting dictionaries, and preserving the resulting transliteration and translation.
When used in the browser, project data can remain stored locally, keeping manuscript images and working files private. Projects can also be backed up through Google Drive, while transliterations and translations can be linked with Google Docs for bulk editing outside the Şerḥ interface and synchronized back into the project.
Şerḥ does more than simplify the transliteration workflow. By linking each transliteration and translation to its precise geometric location on the manuscript page, it also turns ordinary scholarly transliteration into structured training data. Corrected projects can therefore serve as ground truth for a researcher’s own line-segmentation, handwritten text recognition, and keyword-spotting models.
Preparing a manuscript
Şerḥ accepts page images and PDFs and turns them into editable manuscript projects. Before transliteration, each page can be divided into reading areas and individual text lines. These line regions determine reading order and allow the transliteration editor to keep the relevant portion of the manuscript in view while you work.
Line segmentation can be drawn manually, imported from existing YOLO output, or proposed by an experimental in-browser model or an external vision model such as ChatGPT, Claude, or Gemini. Whatever method is used, the result remains ordinary editable geometry: lines can be moved, reshaped, split, merged, reordered, or deleted before the page is approved.
Transliteration, normalization, and translation
Once a page has been segmented, Şerḥ provides a line-by-line transliteration workspace. Selecting a line on the manuscript activates the corresponding text field, keeping the image and transliteration aligned as the document is read. The diplomatic transliteration can be entered in Latin script, Ottoman Arabic script, or as a deliberately incomplete reading when the text remains uncertain.
A separate normalization layer allows a diplomatic reading to be expanded or regularized without overwriting it. Translation is likewise kept alongside the transliteration, with space for literal, readable, and user-edited versions. This makes it possible to preserve the distinction between what can actually be read on the page and later editorial interpretation.
Dictionary lookup in context
Words in the active transliteration can be sent directly to Ottoman lexical resources without leaving the manuscript workspace. Şerḥ performs local lemmatization and conservative suffix analysis for searches in Osmanlıca Sözlükler, while reconstructed Ottoman spellings can also be used for more refined dictionary searches and, where available, LexiQamus.
This is particularly useful for forms that are difficult to identify from a diplomatic transliteration alone: the transliteration, reconstructed Ottoman spelling, dictionary search, and manuscript image remain visible as parts of the same reading process.
Desktop focus editor
The optional Şerḥ desktop app adds a focus editor designed for sustained manuscript reading. It keeps the line editor, AI chat tools, and dictionary interfaces together in a single window, reducing the need to move between separate applications while working through a difficult passage.
The desktop app also handles keyboard input switching automatically. Clicking into LexiQamus can switch input to an Arabic QWERTY layout for Ottoman-script searches, while returning to the transliteration editor restores the Latin-script keyboard without requiring the user to change input sources manually.
Dictionary searching is also linked across tools. A form identified or searched in LexiQamus can be carried directly into Kubbealtı Lugatı, Osmanlıca Sözlükler, and Keşf-i Yeni Redhouse, making it easier to compare lexical evidence across several reference works without repeatedly retyping the same word. Together, these features make the desktop app particularly useful for close, line-by-line reading, where frequent movement between transliteration, dictionaries, and AI-assisted interpretation would otherwise interrupt the flow of work.
Optional AI assistance
AI features are optional and are designed as editorial aids rather than replacements for scholarly review. Depending on the configuration, a model can propose line segmentation, suggest normalized readings, check a page, or generate draft translations. Suggestions remain editable and can be accepted, rejected, or ignored.
Şerḥ can work with hosted API models, or prepare bounded review tasks for local coding assistants such as Codex, Gemini CLI, or Claude Code. The experimental Şerḥ line-segmentation model runs entirely in the browser and does not require an API key.
AI-assisted review can also function as a kind of spell check for scholarly transliteration. Rather than silently rewriting the text, Şerḥ can flag apparent inconsistencies in diacritics, transliteration conventions, morphology, or readings that appear unusual in context, allowing the researcher to inspect each suggestion against the manuscript itself. This is particularly useful in long transliterations, where small inconsistencies in forms such as ḳ/k, ż/ẓ/z, vowel length, or editorial conventions can otherwise be difficult to spot.
Automated segmentation, reconstructed spellings, transliteration suggestions, and translations remain proposals for scholarly review. They are intended to accelerate comparison and checking, not to represent uncertain readings as settled text.
Saving and reusing the work
Transliterations, normalized text, translations, notes, and editorial decisions are saved as the project develops. A complete project can be exported as a .serh backup containing both page images and editorial state, allowing the manuscript to be restored later or moved between devices. Projects can also be backed up through Google Drive.
For longer-form editing, transliteration and translation fields can be linked to Google Docs. This allows the same text to be edited either within Şerḥ, alongside the manuscript image, or in a conventional document interface better suited to sustained revision, commenting, or bulk changes. Edits can then be brought back into the Şerḥ project without severing the connection between text and manuscript geometry.
Corrected segmentation and transliteration data can also be exported for reuse in HTR, line-segmentation, or keyword-spotting training, making the work produced during ordinary transliteration available for later computational research.