# tacotron-2-support
- e
Eugene
10/21/2022, 11:15 PMAight - r
Reclezon
10/21/2022, 11:17 PMPhones about dead so brb ig - e
Eugene
10/21/2022, 11:17 PMThank you guys for helping me :) - r
Reclezon
10/21/2022, 11:20 PMDoes sound a bit grainy for background noise but the voice is very good otherwise. - e
Eugene
10/21/2022, 11:24 PMIs there anything to conclude from these images? - u
{K EY1} (Kei)
10/21/2022, 11:27 PMThat means the model is very good - u
{K EY1} (Kei)
10/21/2022, 11:27 PMThe first thing is a map of the generated phonetics One on the right is an alignment graph (straight orange diagonal line is good) Emojis are just whatever is detected in the text - e
Eugene
10/21/2022, 11:29 PMHmm - e
Eugene
10/21/2022, 11:30 PMI do hope that with more voice lines the model could become more realistic. - e
Eugene
10/21/2022, 11:30 PMAnd I also find quite some words in the output audios hard to understand. - r
Reclezon
10/21/2022, 11:42 PMMore is always better, though mind that rule about exponential decreases or whatever. - e
Eugene
10/21/2022, 11:46 PMWhat's that rule? - l
legorunnerkid
10/21/2022, 11:55 PMI remember having my model sound very weird when I thought I had it. - l
legorunnerkid
10/21/2022, 11:55 PMI believe it had something to do with me not waiting enough. - r
Reclezon
10/22/2022, 12:16 AMLaw of diminishing returns! I remember now - l
legorunnerkid
10/22/2022, 12:53 AMWhat's interesting about the source I'm using is that certain cutscenes with echos have two audio channels. - l
legorunnerkid
10/22/2022, 1:25 AMI finally got 200 voiceclips transcribed now. - l
legorunnerkid
10/22/2022, 1:29 AMI've got about 3 minutes of voicelines. - u
🏝✍Berkeley•Ray•Brewski
10/22/2022, 1:48 AMAs someone who went through 165; I can't help but be proud of you! - l
legorunnerkid
10/22/2022, 1:49 AMOh wait, I was meant to post this in #768215837248716819. - l
legorunnerkid
10/22/2022, 1:49 AMSince I technically don't need help. - g
Gamma Prime
10/22/2022, 3:48 AMDoes setting a batch size of 1 hurt models? - g
Gamma Prime
10/22/2022, 3:48 AMI did that on accident and didn't realize it until training started. 50 epochs in, the model was stroking pretty bad. - g
Gosmokeless28
10/22/2022, 5:05 AMYes - g
Gamma Prime
10/22/2022, 5:10 AMI had a few samples in the dataset that might not have been doing him favors either. Turns out that Adobe Shasta maybe shouldn't be used to clean Transformer voices. - e
Eugene
10/22/2022, 2:02 PMIs it hard to make a model that works with reference audio (instead of text to speech)? - g
Gosmokeless28
10/22/2022, 2:25 PMNo - a
AhmadGT
10/22/2022, 3:14 PMis it normal that colab delete everything when disconnect? - a
AhmadGT
10/22/2022, 3:25 PMalso is the pipeline better or the legacy? - r
Reclezon
10/22/2022, 3:41 PMPipeline