Slobberbone, le 22/05/2014 à 10:26
While the lib truly is in its very early stage of development, it is already functional : out of 500 speaker voices extracted from Ted.com talks, Recognito identifies them all.

DISCLAIMER : the above doesn't mean anything for real life scenarios, please read on :-)

Indeed, the Ted based test is quite biased :

   overall good quality recordings
   professional speakers usually speak loud and clear
   vocal samples (both training and identifying ones) were extracted from a single recording session, which means the surrounding noise and the average volume of the voice remains stable

So the voice print extraction works as advertised but "probably" won't be able to cope with vocal samples of the same speaker coming from various recording systems with huge differences and/or very different sounding environments.

Please note, I used the word "probably" so basic testing on your side should rapidly provide better insight on whether or not the current state of Recognito is suitable for your particular use case.

There are already people out there quite satisfied with the results... I'm a perfectionist aiming for state-of-the-art technology :-)
(Permalink)