This allows you to perform speaker verification using Overflow 
(http://freespeech.sourceforge.net/overflow.html). Note that you need at 
least version 0.6.0 for this to work.  You need to compile and install 
Overflow using the directions on the web page.

You need to create your voice GMM (gaussian mixture model). To do that,
you need to record ~1-2 minutes of your voice at 16 kHz using 16 bits/sample (little-endian) mono and with no header. If all you have is a .wav, do:
sox my_voice.wav -t sw -r 16000 -c 1 my_voice.sw
To check if the file is OK, do:
play -r 16000 -t sw -c 1 my_voice.sw

Go in the directory where you find train_sd.n and si.dgmm and do:
batchflow train_sd.n my_voice.sw my_voice.dgmm

batchflow is part of the Overflow package, make sure it is in your path.


Now, the interesting part: speaker verification. Once again, you need a file in the same format and do:
batchflow score_sv.n some_file.sw my_voice.dgmm

if some_file.sw is in fact your voice, the output result should be positive. 
Oterwise, it should be negative. In case you want to play with it, I have 
also included a model of my voice: jm.dgmm

Bear in mind that this is the first speaker verification prototype and that 
accuracy will likely improve in future versions.


If you would like to see how everything works, you can edit the "program files"
score_sv.n train_sd.n with the "vflow" application (also provided with 
Overflow). If you want to play with training of the speaker-independent model
(si.dgmm), you'll need to have your own training data.

Have fun!

	Jean-Marc Valin (valj01@gel.usherb.ca)
