Yes you can, in theory, solve without training data it can be solved as a graph problem also.

[knasim]

#1. Wish we can add comments in github main project page.

To what you said: "I am pretty sure it is possible to implement a prediction approach that does not require to collect training data at all. Given that the user types a text in a certain known language (e.g. English), the statistical information about N-grams in that language, combined with the similarity metric of the detected keypresses could be enough to detect the text being typed. Effectively, it boils down to breaking a substitution cypher."

My Answer: I could entirely wrong, I usually am. But yes you can solve as I stated earlier as graph problem.
Use case: Given an arbitrary key, a graph of nearby keys can be constructed for each key on the keyboard.
From these graphs, you can analyze the waveforms and developing a ranking to predict which key was pressed from the peaks.
And then store the waveform rank back into the graph making it a weighted graph and tells you which key was pressed.
Ofcourse applying some sophisticated graph techniques. It becomes a combinatorics problem also.
This could improve the "nearby key" approach.

This is a very interesting problem to solve - and the accuracy could be nearly perfected. I wonder if google listens in to our keystrokes to improve their AD algorithms.

From the graphs you can create better training data too ?

[ggerganov]

Thanks for the ideas and suggestions. You might want to read my recent post on this topic in the "Discussions" section:

#31