I had recurring problems with Leo ‘transcribing’ blank pages. It coped amazingly well with pages of writing which had all manner of illegible handwriting, ink spots, strike outs, and even where ink had bled through from the other side. However, it had trouble recognising blank pages. This may seem like it shouldn’t be a problem but it took some time to then remove the transcription of random numbers, tables or phrases in order to export. On one blank page it added a huge table of 1-800. This possibly happened because the scan showed a section the next page in some cases, or was showing ink from the reverse?
This is helpful feedback. Previously we had an issue with there being too many blank pages in the training data, such that Leo often gave empty transcriptions when there was text on the page. But we could try to add some more examples of real blank pages to the training data so that it knows when to do this. Thank you!