ill try my best

Since you guys were finding out how to use reference audio with tacotron 2, watch this video and it’s lip sync, it’s not pitch corrected (I don’t think) but in the description they say they use tacotron 2 AND WaveRRN models so WaveRRN could help:

I know but what I meant is it’s not live like deepfake live

wavernn runs after tacotron, it doesn't affect the rhythm either the audio or video was manually edited to sync up

Well what is WaveRNN used for?

it turns the tacotron output into audio in a smart neural™ way

we use hifi-gan for that which isn't perfect but is fast

When you say that do you mean to make it have more emotion or make it not sound robotic?

the latter

if you've ever trained a pipeline model the preview audio there sounds warbly

that's because it's not using hifi-gan

That's why? then what does the pipeline notebook use then? a custom thing?

i mean the training notebook

the preview audio there

the synthesis notebook does use hifi-gan

THats what I thought you meant?

I was wondering why it sounded way different than the synthesis nb

what are the best voices on uberduck? (best quality score)

is there a way to see them?

thanks

actually how do i do that

two ways

1: in the transcript, put

|

and then a speaker number after every line

if you have 4 speakers then they'll be

|0

,

|1

,

|2

,

|3

and whats the other way

2: split it into multiple datasets

and let the notebook merge them