Transcription demo (single image only)
Upload or drag-and-drop an image of any historical document (including tables of data) to transcribe it to text.
Please note:
- As with any AI-assisted system, there can be errors in the output - human verification is recommended.
- This website is under active development and is offered here as a demonstration only. You may see errors, or things may stop working altogther at any time!
- Please email me with your feedback, comments, observations, suggestions, etc.
How does it work?
I use a two-model approach to transcription.
- First the text on the image is recognised using Amazon's Textract. This does a great job of picking up printed text and a so-so job with handwriting.
- Next the OCR from Textract and the original image are both provided to a Google Gemini model for transcription into 'Markdown' format which simply allows the model to mark parts of the text as headings, bold, italic, etc, and also present information in tables.
- Finally the image and Markdown transcription is presented to you.
I do not claim that there is anything groundbreaking about this approach, but it seems to produce notably better output on historical documents than other transcription services currently available (eg, HandwritingOCR, Pen2Txt, NanoNets, etc.