Ah sure yep

there are not enough voices in uberduck that support arpabet it must real

Pretty quick as I recall, had a go on their notebook a while ago

Audio samples: https://nv-adlr.github.io/Flowtron

Oh seriously? Maybe I should stop trying to revamp old code and just use this

I want something with inference of ~10secs or so on my old 1660 gpu w/ cuda support + fp16 optimization, gonna turn it into a tts bot for voicechats

SV2TTS had passible inference time from my testing locally

But it's a lil outdated and I figured theres better models now

I mean Tacotron sounds just as good if not better at times when using Hifigan. But yea the style transfer and ways you can modify speech defo seem cool.

Maybe theres a way to add Hifigan support to SV2TTS

Gotta actually understand ML tho for that

Vs mash Lego block code together

I have no idea

Also SV2TTS is using tacotran1, dunno how much improvement is between tacotran1 and tacotran2

Probs quite a big one, considering the speed of the tech.

Emotive text is also nice

Absolutely

If you use some sorta NLP model to label your data

I'll dump any datasets I create while messing with my stuff here as thanks for this discord providing such great resources

Yea surely.

Speaking of dataset has anyone dealt with the MLP dataset?

I was gonna mess with it

Seems super high quality

Would be great for training a multispeaker model

I have a question about formatting

For the MLP dataset I know I have to construct the dataset in the following format: https://raw.githubusercontent.com/NVIDIA/tacotron2/master/filelists/ljs_audio_text_val_filelist.txt Yet i know that the MLP files are in the following format:

2.202059    3.932059    00_00_02_Pinkie_Neutral_Very Noisy_To the right! to the right!
3.932059    5.672059    00_00_04_Pinkie_Neutral_Very Noisy_To the left! to the left!
5.672059    7.202059    00_00_06_Pinkie_Neutral_Very Noisy_Now up! now up!

Im a noob so don't know if theres significance. But I've made models properly laugh when I label laughs as "haha", so long as dictionaries arent being used ofc. So I dont see why you couldnt expand that.

Do I have to parse it so that it's in the correct format as in converting

2.202059    3.932059    00_00_02_Pinkie_Neutral_Very Noisy_To the right! to the right!

-> DUMMY/Pinkie.flac|To the right! to the right!

i guess

also that format looks like it was exported straight from audacity