So no one is gonna help

if you don't show the error nobody can help you lol

RuntimeError Traceback (most recent call last) in () 4 hparams.gate_threshold = 0.1 # Model must be 90% sure the clip is over before ending generation (the higher this number is, the more likely that the AI will keep generating until it reaches the Max Decoder Steps) 5 model = Tacotron2(hparams) ----> 6 model.load_state_dict(torch.load(tacotron2_pretrained_model)['state_dict']) 7 _ = model.cuda().eval().half() 1 frames /usr/local/lib/python3.7/dist-packages/torch/serialization.py in __init__(self, name_or_buffer) 240 class _open_zipfile_reader(_opener): 241 def __init__(self, name_or_buffer) -> None: --> 242 super(_open_zipfile_reader, self).__init__(torch._C.PyTorchFileReader(name_or_buffer)) 243 244 RuntimeError: [enforce fail at inline_container.cc:145] . PytorchStreamReader failed reading zip archive: failed finding central directory

looks like you didn't provide a valid model file

or the model file is broken

that's the broken file?

Yea?

then you should fix it

This might sound terrible at the moment cause it's still training...oh btw I suck at gathering dataset lol

Sounds good already.

How long do I have to wait

For ever 😈

say sike right now

No 😈

around 40 epochs should be fine

dammit, it disappeared

how?

I zoomed in

So I was wondering, has anyone here used Jukebox or other similar VAQ-VAE systems for speech synthesis purposes?

I did a bunch of experiments with it last year and one of them was continuing a piece of spoken word audio, from a character with a very distinctive voice who was almost certainly not represented in Jukebox's training data - and while the resulting continuation wasn't really legible compared to the input text, it recreated the character's distinctive prosody spot on

Could be useful for auto tuning maybe

Fifteen AI had something need where you could make a character speak a certain way but making it say another sentence first and then having the AI read the second sentence in the same way it would have said the first sentence

Something like this might be useful for “guiding” the AI

emotional contextualizer with deepmoji i think

We were talking about that earlier but we need a way to implement it. I’ve been working on a data base to make emojis for certain things but I’d much rather go through deep emoji make some corrections and use that index instead

But I have no clue what I’m doing beyond pretty pictures so talk to Pi about it

Dark thoughts of a cold dystopian future but I wonder if you could use facial recognition software on a TV show and have it learn their expressions so it can give your emotions as time stamps that would make for easier transcription allowing for implementation of emotions into their transcriptions

emotion from voice is used in a lot of call centers already