all

so set the thing to

["null"]

it'll get better

hopefully

btw i cut the silent parts with this script https://raw.githubusercontent.com/Yeongtae/tacotron2/master/preprocess_audio.py

iiinteresting

i personally use ffmpeg, though the version in colab doesn't have the silence removal filter so i have to download it from elsewhere

Can anyone either explain to me or point me at a resource that gives an ELI12 version of the key config parameters for training? I'm trying to understand things like: How to decide what the learning rate should be based on dataset size when finetuning?

honestly idk if there's much of a science to it just if you have a very small dataset then decrease batch size (as low as 1 even) and learning rate (no lower than 1e-4) with anything above 10 minutes the default settings should be good (though lowering batch size is apparently always a good idea if you've got time)

WHAT IS BRO DOING ‼️ 📢

why does this keep happening

WHAT IS BRO DOING ‼️ 📢

why is the val loss a rollercoaster

Ur not decaying lr

crust does not do that

Hecko did not implement it

what is the course of action to implementing lr decay (i would expect it to be part of pipeline config)

Look inside the legacy colab training

Scroll down to the training loop

how does one implement it

i got info about a reduction window schedule yesterday but that's something else

Reduction window speeds up the learning (not learning rate) per step

> but that's something else

adam optimizer

Cyber Quincy test:

what happening?, this is gonna better error is help

@tylerdurdenceketi wait how do you set the symbol set to ipa in pipeline

/usr/local/lib/python3.7/dist-packages/uberduck_ml_dev/text

What did you edit?