I don't trust Speech to Text

1. 2017-11-24 at 8:15 PM UTC

   #21
   

   bling bling Dark Matter
   throat michrophone implant in small infant babbys
2. 2017-11-25 at 7:16 PM UTC

   #22
   

   Lanny Bird of Courage

   Originally posted by jedi.Goldstein You are right to be.
   PLA Episode 17 was about voice password hacking:
   https://www.youtube.com/watch?v=tw8L95hrxlE
   
   There are also tensor based methods emerging to clone human speech patterns (Lyrebird).
   https://www.theverge.com/2017/4/24/15406882/ai-voice-synthesis-copy-human-speech-lyrebird

   
   The NN claim is a lie, maybe prosodic estimation is done by NN but synthesis itself is clearly carried out by MLSA. It's a mistake either way as unit selection would be a more reasonable strategy for this use case.
3. 2017-11-25 at 9:11 PM UTC

   #23
   The Self Taught Man Black Hole

   Originally posted by Lanny The NN claim is a lie, maybe prosodic estimation is done by NN but synthesis itself is clearly carried out by MLSA. It's a mistake either way as unit selection would be a more reasonable strategy for this use case.

   
   I tried out Lyrebird's online beta, and the result sounded like a nasally Microsoft Bob.
   
   Adobe have a similar product in the works, demoed it last year.
   
   Post last edited by jedi.Goldstein at 2017-11-25T21:20:59.973663+00:00