Bigrams across uncertain spaces

[MarcoP]

I feel like we cannot ignore the role of the transliterator in this context. 

I agree: Emma pointed out an important side of the  subject. Of course it's easy to imagine that a transcriber transliterator is influenced by word structure. 
Another missing bit is how spaces work in readable manuscripts: we know that there is a great variability in how spaces were written, e.g. prepositions were often attached to the following word. But I don't think any transcription handles something like "uncertain spaces".
This is an area where one might think of using image processing: e.g. the number of spaces in each line (N) is extracted from a manual transcription and given as input to the software; then the N wider spaces are classified as certain / uncertain according to a threshold on a computed measure.