Uberduck

Yeah i had good results with the arpabet legacy notebook too but eh
I think the current recommended notebooks are in <#841437191073955920>

Is there any alternatives to like an audio converter that I can use to make my zip files that are similar to this that will work on mobile specifically.

how long do i train on pipeline for 50 seconds of audio

when you train on legacy, is it just supposed to...stop on its own?

When it reachs the epoch limit it justs stops

You need to change the epoch limit higher

But unless you have a gigaton of data, there's no need to go past around 150~ epochs, it won't make the model better

i have 24 clips, how many epochs sould i use?

set it to 75 and it finished in 14 minutes 💀 💀💀💀

i got this in the polish pipeline notebook...

same error as the one i got here training on the current pipeline notebook with the polish base model

all i can think of is some transcript issue somewhere

but unfortunately pipeline shuffles the wav·s so there's no info on where exactly to look

i see... i'll look through the smaller speaker's transcript
what sort of issue could it be?

run it through this https://www.browserling.com/tools/letter-frequency

```a: 1618
.: 785
w: 730
v: 730
s: 702
o: 639
0: 551
i: 544
e: 390
n: 370
/: 365
|: 365
k: 365
↵: 364
d: 266
u: 238
m: 208
r: 194
t: 193
1: 185
2: 167
3: 128
g: 91
ł: 85
j: 85
4: 76
5: 73
6: 73
z: 73
c: 70
7: 65
8: 63
9: 61
h: 48
ś: 39
b: 39
A: 18
f: 11
p: 11
ć: 2```

ran the other transcription through too, i dont see anything weird

wait is the fact that the encoding of the text file is in Windows (CRLF) a problem?

it shouldn't be an issue, i do crlf all the time

in notepad++, replace `.*\|` with nothing (with regexes enabled) and then sort by length