# machine-learning
- t
tylerdurdenceketi
06/21/2022, 4:49 PMsymbols.py - u
{K EY1} (Kei)
06/21/2022, 4:49 PMAh - u
{K EY1} (Kei)
06/21/2022, 4:49 PMTy - t
tylerdurdenceketi
06/21/2022, 4:49 PMis this ok - t
tylerdurdenceketi
06/21/2022, 4:50 PMusing crust as base for turkish - u
{K EY1} (Kei)
06/21/2022, 4:50 PMYeah, it'll clear up eventually It looks like it's clearing up - t
tylerdurdenceketi
06/21/2022, 4:50 PMsometimes it gets messed up - u
{K EY1} (Kei)
06/21/2022, 4:51 PMIt's probably pretty early on in the training still If it still looks like that later, there could be a problem, but it should be fine - u
{K EY1} (Kei)
06/21/2022, 4:51 PMYou have 5 speakers? - t
tylerdurdenceketi
06/21/2022, 4:51 PM10 - t
tylerdurdenceketi
06/21/2022, 4:51 PMxD - u
{K EY1} (Kei)
06/21/2022, 4:51 PMOH - u
{K EY1} (Kei)
06/21/2022, 4:51 PMI see - u
{K EY1} (Kei)
06/21/2022, 4:51 PMI haven't done custom language on pipeline yet, other than a failed attempt at japanese - u
{K EY1} (Kei)
06/21/2022, 4:53 PMI prefer legacy, it's less confusing At least pipeline can do multispeaker though, which is incredibly useful - t
tylerdurdenceketi
06/21/2022, 4:56 PMwhat if i use plain text instead of ipa? - t
tylerdurdenceketi
06/21/2022, 4:56 PMi want to believe it will converge at the end - c
Couch
06/21/2022, 4:56 PMif plain text is used you would have a character-based turkish model rather than ipa phones - t
tylerdurdenceketi
06/21/2022, 4:57 PMwhat are pros and cons - c
Couch
06/21/2022, 4:58 PMplaintext gives you less control over pronunciation whereas in ipa you give it the exact phones for the model to pronounce (same for arpabet except with english phonemes) and also ipa is multilingual because International Phonetic Alphabet - t
tylerdurdenceketi
06/21/2022, 4:58 PMand may i ask a question what will happen if i don't cleanup silences at the start and end of audio files - t
tylerdurdenceketi
06/21/2022, 4:59 PMthat sounds legit - u
{K EY1} (Kei)
06/21/2022, 4:59 PMIt might have trouble converging I had that issue with a model; it learned really weirdly and never got to proper speech - c
Couch
06/21/2022, 4:59 PMand also silence at the begnning and end of the audio would cause the alignment and the audio to misalign - u
{K EY1} (Kei)
06/21/2022, 5:01 PMThe other alternative is it might just learn to speak with silence at the beginning and end But it might not - u
{K EY1} (Kei)
06/21/2022, 5:02 PMFor me it's not worth the risk so i always cut silence (i do it manually though) All the research corpuses i've used haven't had that problem though - c
Couch
06/21/2022, 5:02 PMdont really know how to visualize this but - t
tylerdurdenceketi
06/21/2022, 5:03 PMbtw i have a model tracey from gta v I uploaded it to the uberduck but it was not approved - t
tylerdurdenceketi
06/21/2022, 5:07 PMwhich layers should i ignore when i train from crust? and which layer does what? - c
Couch
06/21/2022, 5:07 PMsomeone in the annoying orange fan club is attempting to train spanish crust and i think these were the layers to ignore["speaker_embedding.weight", "spkr_lin.weight", "spkr_lin.bias", "embedding.weight"]