theres a neat little python doodad that can take care of arpa conversion

....and I may g4et the others lol

how do i get rid of the chirping sound, I am not really good at editing audio

or do i even need to edit it out?

audo.ai

should work well

for sfx

For submitting a voice (first time) Do i make a file (.zip) for them to open with a bunch of audios?

also what model type would it be for text to speach (Gandalf)

WAVs must be Mono, 22050 Hz, 16-bit and in a folder named wavs. They must not contain any background music or sound effects, if they do, try using websites like mvsep.com for music and audo.ai/noise-removal for sounds or if it's a show/movie, get a 5.1 audio track if you can which doesn't have music but some sound effects will still be there. (I recommend exporting the voice in the highest audio quality possible, isolating it and then converting it to the supported format.) You can use an audio editor like Audacity to gather audio. Then you make a txt file, which goes like this wavs/1.wav|Your text here. wavs/2.wav|Your text here. wavs/3.wav|Your text here. And so on. Commas, periods, exclamation points and question marks are supported. You can also add ARPAbet support to your model by transcribing sentences in ARPAbet. Then you train your model using a notebook in #841437191073955920. (Your voice folder should have a txt file and a folder containing WAV files named wavs.)

submitting a dataset is not akin to submitting a model

a dataset will not result in anything showing up to the website

in fact i dont think anything is done with datasets

For example, this is some text for my LGR set I posted a test of above.

wavs/LGR10.wav|It might look like a standard digital organizer first glance but once you're booted up and poke around it's apparent this thing is a legit palmtop computer. wavs/LGR12.wav|Inside is an Intel ATC 88 a processor running at 4.92 Megahertz and 128 whole Kilobytes of RAM. wavs/LGR14.wav|And it's also got a proper dos operating system capable of running PC software either stored internally or on memory cards. wavs/LGR16.wav|Right alongside the built-in applications. It even acts as a quick ATM hacking tool when the mood strikes. wavs/LGR18.wav|Or at least it does in Terminator 2 if you're young John Connor. wavs/LGR110.wav|appearance in that movie being the portfolio's most famous claim in real life though the po fo as it was known to its' fans. wavs/LGR112.wav|was a machine released in 1989. wavs/LGR114.wav|being revealed at that year's summer Consumer Electronics Show in Chicago. wavs/LGR116.wav|And it had a suggested price of just $399.95. wavs/LGR118.wav|when it hit store shelves making it surprisingly low cost compared to the competition. wavs/LGR120.wav|not that there was much competition since the entire portfolio was the only thing like it on the market. wavs/LGR122.wav|for a short time being credited as the world's first palmtop computer. wavs/LGR126.wav|first though I want to give a huge thanks to L G R viewer Brent. wavs/LGR128.wav|for donating this particular po fo along with letting me borrow a pile of add-ons peripherals and software.

>4.92

oh yikes i wouldn’t recommend doing that

if he says it like four point ninety two just spell it all out

idk if the cleaners would interpret it correctly

They should be put in an archive or something so people can retrain or edit them as needed

https://www.speechtexter.com/ - I also find this site extremely useful, so you can spare your poor keyboard some pounding.

And good training will look like this, this is from the new LGR model currently in the oven.

wavs/Sin1.wav|I AM SINISTAR. BEWARE, I LIVE. wavs/Sin2.wav|RUN, RUN, RUN! BEWARE, COWARD. wavs/Sin3.wav|RUN, COWARD! I HUNGER. - no clue if this will work but I'm trying it anyway.

there's still no hifigan finetuning for pipeline models, right?

or does the hifigan finetuning notebook work on those models

mega b gave me a nb that supposedly allows for it

and warmstarts from a crust trained multispeaker hifi

too large batch size

↑ to elaborate: if your batch size is 14 and your dataset is 10 wavs, it'll make 0 batches because it won't fill up a batch and so won't train anything and you just get the base set it to at most your dataset's size, but preferably something even lower (as low as 1) with a very low learning rate also don't you dare insult linda johnson