We are happy to introduce a new PyLaia print model (Transkribus print 0.3). You may already be familiar with our HTR+ print model, which in addition to common Antiqua and Fraktur typefaces can also decipher typewritten text, modern computer printouts, and even various unusual ‘decorative fonts’ in several languages. A similar model is now also available for PyLaia.
We have compared the two models and the results of the PyLaia model seem to match and in some cases surpass those of the HTR+ model. On one of our test sets for example, the new PyLaia model was 30% faster while having a CER of 1.28% compared to 1.64% of the HTR+ model. We have observed before that PyLaia seems to be doing very well on large and diverse train sets such as this. Below you can see some results of the new model, but the best way to see what the model is capable of is to just try it out.
Also, for those of you who have to keep an eye on project budgets: HTR processing with Pylaia models uses slightly fewer credits than with HTR+ models.