Learning by storing bad results as future guessing reference

[xtrasmal]

Hi there,

Thanks for this great toy.

Humans can interpretate badly spoken sentences

and can guess the right intentions. The computer should learn how to make right guesses.

To make the speech control work.. I need to become a speech acrobat.

What I do is try different English accents, or Dutch accents.. or even say the English words like if they were Dutch.. just to get the speech computer WORKING.
This get's shitty pretty fast.. so my plan is to stop making an effort and go with the error flow.

Without context, you cant guess

Like when I would say: understand
The result of the speech might be: on the sand
Without context.. nobody would on the sand me ( :] )

But wait.. there is context

Ofcourse we do have context when we build or own app. The app is the context.
Secondly and fortunatly, most apps have limited functionality.
Apps can add posts, send messages, create events.. and maybe a couple of things more.

Limited functionality makes an easy learning place

Let's say that we want to add a new post to our website by speech.
I would say: add a new post
Speechcomputer might return: at ale new pushed
App does: nothing

Because we have limited functionality, we can store the bad result
as a synonym for the right result.
So the next time:
I would say: add a new post
Speech computer might return: at ale new pushed
App does: Adding new post

Plan is to test it some day soon.

1. Setup and API with PHP/SQL/ANGULAR with routes that lead to actions
2. Make up sentence commands that lead to the actions in step 1
3. When a result does not match, my app will ask me what I wanted to do.
   I select the expected action and store the unmatched in a database
   and use it as reference. So the next time, the same mistake will lead to
   a good result action. Maybe doing pattern matching or database find score,
   to let the app 'feel more confident'.
4. After doing step 3 many times, My app will give better results.

So by doing this, I expect to have an app that will 'grow' on it's owner after a while.

What do you think?

[janantala]

Hi,

Thanks for the feedback.

I am also not a native English speaker and have problems with my accent so I know what do you mean.

This is one of the biggest problems of speech recognition. Google server already returns more results with percentage and I use just the first (others are mostly wrong too).

I plan to work on it, but I want to do it client side only (local storage hash map) with no external dependencies.

So far there is regex and you can target these error results and log all new with google analytics or something else.
You can also change recognition language from en-US to another (Dutch).

[xtrasmal]

When I speak Dutch, the recognition goes very well.