> Perhaps speech could be translated to a string of word > alternatives with probabilities, and language > understanding gives a hint which alternatives make more > sense.

This is already done, but probabilities are used to determine the best alternative.

So basically we ask

    if p("which" | "Ding dong", "is dead") < p("witch" | "Ding dong", "is dead") 
    then print("witch")

Or basically, which of the alternatives have the highest probability given the context and we find a way to measure similar contexts.

This generally doesn't give us an understanding of the text.

Yes. I didn't mean to propose that. I know that speech recognition already considers alternatives and probabilities.

The main focus of my proposal is something different: a more intuitive «understanding» using neural networks and lexicons. This is somewhat orthogonal to speech recognition.

Then I suggested that this «understanding» could help make speech recognition more reliable by adjusting of some probabilities.

You see, I have an important insight as a Deaf person. Lipreading only offers a small part of what a person could understand by hearing. But I am often still able to understand what a person says by guessing and using the discourse context to fill in the blanks. I am sure that we all do that to some extent, or we wouldn't have homonyms in the language.

This insight tells me that perhaps a disrupting innovation would be a more natural «understanding» of text (in written form) with the help of neural networks and a lexicon.

Can you give an example of something you want this system to understand?

A «misuse» would be a spying system which automatically «understands» euphemisms for forbidden things.

For me an important use would be a tool to make speech recognition more reliable, for example to automatically subtitle Youtube videos with better accuracy.