1. Help Center
  2. Text Recognition

4. Super Models

We strive to regularly add new elements and new technologies for our to enhance the power of Transkribus. One of these technological elements are the Super Models for text recognition, which are the most advanced models we can offer so far.

Previous step: Public models

Transformer-based text recognition models - or Super Models - are great at dealing with natural language, which helps them decipher all sorts of written materials. They are able to outperform existing PyLaia models by a significant margin.

A key advantage of these models is that they consist of both an optical part that processes the images and an extensive language model that tries to make sense of and improve the extracted text information. The language part is able to deal with more than a single language, and old and new forms of a language at the same time. This allows the models to output very high-quality text.

Thus, the first Super Model we have trained, The Text Titan I, is aimed at tackling the challenges of recognizing both handwritten and printed text with remarkable accuracy and efficiency, in a variety of languages. 

Benefits of using Super Models 

As explained above, a Super Model is one big, very general model with the ability to recognize both handwritten and printed text simultaneously. This will be particularly useful when working with mixed materials.

Some archival holdings or manuscript collections can have different types of writing, printed as well as handwritten documents, pre-printed forms filled in by hand, index cards etc. With Super Models, you are able to use the model on both types of text, which means that you don’t need different models or have to constantly change settings if you are working with handwritten as well as printed documents.

The Text Titan I. is remarkably adept at processing a wide variety of materials and writings. Although our Super Models are not currently fine-tunable or trainable by users, they deliver outstanding out-of-the-box performance across numerous heterogeneous types of material which will help users to quickly produce Ground Truth for training customized PyLaia models.

It is important to note that Super Models cannot be used as base models for custom model training due to their extensive amount of data.

As we recognize the value of customization and speed, we are diligently planning for the future where Super Models like the Text Titan I. will be made adaptable to better meet your specific needs.

A specialized PyLaia model trained for well-defined material can, however, still yield better results, but the creation of the training data for such a specialized model can be sped up considerably by first processing part of the material with a Super Model and correcting it manually.

Next step: Credit system