I messed with some learning rates in talknet and got a pretty good results but the samples where set up different

I think some of the phonemes are a tad messed up in my transcript, I manually converted English Graphemes to Arpabet Phonemes but I could not tell if the changes I was making where helping I am still trying to get my bareing on what phonemes are what in arpabet

also the voice is hella weird so I have to take creative liberties in how I interpret it on what the hell he is saying.

not all the samples are super clean cut on what they are some are in russian but it's like gibberish it does not even sound like he is saying, but I can't afford to ditch the data as in testing it's been crucial to the model getting anywhere from such a limited dataset

You need to cleanup

I was using 150-200 samples and i cleaned it up even it got 159 samples in the end

the samples are probably as clean as they are ever gonna be for what they are

Can you post a sample inference graph?

I will dm

I had a question if someone knows, do I need some of these numbers to differentiate some of these phonemes or are they for some kind of emphasis? like for "opera" {AA1️⃣ P R AH0} I assumed I did, how does it the control the emphasis?

emphasis yeah

1 is primary stress, 2 is secondary stress, 0 is no stress

oh I understand now thanks

hecko do you still have that tkinter dataset validation tool thing

i need it to check if my transcripts still align after finishing the entire multiple labels thing

mi sona ala e ona

(and the app was purple)

-oh this

thank's

i don't recall if this version supports list.txt import though

sowwy

but it's like 100 lines of python you can probably add it in

i am going to addition it

Huh

how ppl feel about offline access in the browser?

to specific voices

That sounds interesting I already make use of offline tacotron and talknet frontends pretty often because of how convenient it is

given tacotron is already slow to synthesize with full colab cpu i dread to think what it'd be like in webassembly

unless by "specific voices" you mean new ones using a less resource-intensive arch which i welcome