Automatic transcription of handwritten manuscripts from historical archives

Use case ID: 006

Author: IT Department, Chamber of Deputies of Italy

Date: 22 May 2024

Objective: 

Produce textual transcriptions of handwritten manuscripts from historical archives.

Actors: 

• Business units
• Manuscripts of historical archives

Prerequisites:

• Manuscripts of historical archives scanned in a digital format (e.g. in .png format) as training data
• A deep learning model trained on the training data

Scenario:

1. The input images, corresponding to pages of historical archives, are passed through the deep learning model.
2. The deep learning model identifies the rows of handwritten text in the manuscript.
3. The deep learning model transcribes each row and produces the corresponding texts.
4. The texts are shown to the user and are stored in .txt format and .pdf format.

Expected results:

• It takes less time to transcribe handwritten manuscripts from historical archives.
• It takes less time to digitize the manuscripts.
• The manuscripts are provided in digital format. 

Potential challenges:

• The deep learning model needs to be robust enough to manage noise in the scanned manuscript pages (damaged paper, ink spots, etc.).
• Avoid overfitting in order to avoid loss of generality. The deep learning model might be overly fitted to the handwriting of the person who wrote the manuscript.

Data requirements:

• Scanned images (e.g. in .png format) of pages of manuscripts from historical archives

Integrations with other systems:

• Software application used to manage historical archives

Success metrics:

• Time needed to transcribe a page
• Time needed to transcribe the whole manuscript
• The number of characters that are transcribed incorrectly
• The number of words that are transcribed incorrectly
• The number of sentences that are transcribed incorrectly