Manuscripts

The module contains manuscripts, pages, transcriptions and entities, HTR models, training and test datasets, and reports. Recognition and training run locally; newer features depend on build and platform.

Pages, blocks and lines

Import images into Multimedia, then use Insert images… in the manuscript. You can select several images, reconnect a page to an image or remove it.

Whole page offers diplomatic transcription, normalised transcription and translation. Page blocks and lines lets you mark text or image blocks and lines. Adjust geometry with handles; double-click a line's horizontal edge to add or remove a bend point. The eye controls visibility; the column heading toggles it in bulk. Edit line transcriptions in the table; Tab moves to the next line.

Page transcription versions

Each page has its own versions. Adding, copying, renaming or removing a version does not affect other pages. Names must be unique within a page; the main version cannot be removed.

The shared toolbar above both modes provides version selection, management and comparison. Page transcription in Properties shows active/default versions and author, status and origin metadata. The active version determines the edited text. The default is the saved choice for reopening and related functions; viewing another version does not change it. Unsaved editor text is saved to the original version before switching.

Line geometry and order are shared by all versions. Deleting a line removes its texts in all versions and requires confirmation when several exist. Replacing existing lines by segmentation or PAGE XML import is blocked on a page with multiple versions to preserve text mapping. Existing lines can still be recognised and edited manually.

Editorial marks, fonts and entities

Diplomatic transcription supports semantic editorial marks with colour highlighting. Tools → Editorial marks… or the editor button opens the set manager. The built-in set is read-only; a new or duplicated set can change notation, appearance and shortcuts for existing meanings. Sets use JSON import/export. Shortcuts work when the diplomatic editor has focus. Select Vestigia — default to restore the default appearance.

A set changes presentation, not diplomatic text. Its manuscript copy survives removal from the library. Configure editor fonts in Settings → Fonts → Manuscript transcription; confirming updates open editors without changing project text.

Mark a transcription span as Person, Place or Date. The entity catalogue groups occurrences; double-click to navigate to one. Replacing text through recognition can remove entities attached to the previous text; read the warning.

Kraken recognition

Use Tools → HTR models… to prepare an isolated runtime and models. Page recognition distinguishes segmentation and recognition models. Segment text lines before recognition defaults off for pages with existing lines and on otherwise. Without segmentation, polygons remain and results go to a model transcription version.

When segmenting text blocks, areas outside them are masked. Recognise an individual line through its context menu. HTR / OCR retains processing information. Kraken model comparison uses selected lines with known transcriptions.

Party recognition

Party is available in the current development build for Apple Silicon, on CPU. Intel does not support this feature; existing Kraken features remain separate.

1. In Tools → HTR models, select Party (CPU) and Install or repair. Python 3.12 for Apple Silicon is required.
2. Download a Party model from the catalogue.
3. Prepare lines manually or with the existing line detection.
4. Start recognition of one line, selected lines or the page and choose Party.

Party uses existing lines without changing their geometry. The whole batch goes to a Party version as one Undo step; error or cancellation writes no partial result. A compatible custom runtime can be selected and verified in Settings → HTR. Independent model comparison is currently Kraken-only.

Training and test datasets

A dataset accepts transcribed lines from Manuscripts, whole pages from Multimedia, or imported crops and PAGE XML. Use Data check for individual lines and Sources and transcriptions for whole pages. Checkboxes determine the lines included in training or evaluation.

Before training, run Check data quality…. Findings include missing transcriptions, invalid crops, exact image duplicates, different transcriptions of the same image, and problematic characters or spaces. The check does not correct text automatically. Click a finding to select its line; expand character frequencies to see occurrences.

Training and evaluation

Train recognition model… starts local Kraken training from scratch or fine-tuning. Set the base model, epoch limit and advanced parameters as needed. The run stores the data split, epochs, validation CER, charts, full log and model. Splitting training/validation requires at least two usable lines. Register a successful model to make it available for recognition.

Party supports fine-tuning on Apple Silicon/CPU using a Party model and transcribed lines from at least two source images. Training and validation are separated by whole pages. Rectangles are the default geometry. Run history retains validation loss and the run archive; Register model makes the finished model available. Party validation loss is neither percentage accuracy nor CER. Resuming an interrupted checkpoint is not available in the UI.

For an independent Kraken check, use a test dataset and evaluation of one or two models. The test dataset is not part of the training split. Results retain line-level error analysis. Reports provide dataset, training and evaluation templates; export also supports PNG + GT and PAGE XML + PNG.