Francesco
Francesco
Hi, this is taken from the tacotron paper. I believe it helps with "highlighting" the position information for the autoregressive predictions.
Hi, in my experiments the encoder alignments are rather optional, that's why I set it to a lower number of steps than the decoder. You probably can safely set it...
@luis-vera is this solved?
Hi, any further debug information? Does it always occur with the same samples? How do these samples look like?
Hi, are you still having this issue?
I did not try yet, sorry. In case you do, would be great if you reported here.
Hi, r is the reduction factor, it's a technique used to build up attention, which otherwise requires careful tuning and or other less effective techniques. The idea behind is that...
Hi, I struggled to find any literature on this myself, you should find some in the Tacotron paper. All the rest, like scheduling, are rules of thumb I'm afraid.
Hi, that's great! The collapse is a know issue of the autoregressive model. I'm not sure about cover field, are you using convolutions or dense layers after the attention mechanism?...
What is voice control?