QUICKSTART.txt

First off, this program needs good quality audio to work with, so use it 
with a worn microphone preferably. When using it with the 'Google AIY' kit,
try staying close to the microphone and ensure there isn't too much background
noise for best results.

1) type 'make' to compile the program.
2) type './vk' to run it.
3) say a few letter... 'a', 'b' etc... the program should say something like:

BEST=0  or ?, distance=9999999827968, size=20

this is because at this point it doesn't know any letters. 
say 'a', then after the BEST= line appears, type 'a' on the keyboard. 
It will say something like:

*** TRAINED 1 or 'a' size=26

This adds your 'a' sound into the codebook. (a file called cb.txt)
Now when you say 'a' again it should say:

BEST=0  or a, distance=491903, size=24

Now say 'b'. The program should recognize it as an 'a' since this is the only
sound it knows so far.  Add your 'b' sound by typing 'b' on the keyboard.

At this stage you should have a recognizer that can distinguish between 'a' and 'b' being spoken.
Now go through the rest of the alphabet adding an example of each letter in
by typing it after you have said it.

Now you will have a recognizer that should be able to distinguish
between spoken letters with around 50-75% accuracy.

The next stage is to improve the accuracy.
Do this 'correcting' the mistakes it makes by typing the correct letter.
This adds more examples of 'difficult' letters and improves
the accuracy by accounting for subtle variations in pronounciation.

At any time, you can press the 'Escape' key to quit and get a rough
estimate of the perceived accuracy (basically the number of letters picked up 
verses the amount of new examples added.)

Mistakes in training can be fixed by manually editing the 'cb.txt' file. 
The format of this is a letter, then the number of frames that make it up 
(usually around 25), and that many lines of frequency levels. 
For example, you could delete a training mistake by exiting the program and
removing the last chunk of lines if necessary.

Finally, I'd recommend against spending too much time training the program. 
It is essentially a very short and simple 'toy' recognizer,
it is always going to be less than perfect,
but certainly accuracy of 95% or above should be possible.
(especially when used with a high quality worn headset in a quiet setting)