Maybe add a 2x or even quicker to the song player.
Or even make that if a song is for example 5 minutes, make 5 checkpoints, one when it starts, and a few in middle and user can listen to 5 seconds and if the words are synced then it can be approved.
I’ll add it but really therecould be random messups. Models are mostly taking care of it but a little human checks really help train.
Wow!
Can this be added to podcasts?
Ask me in a few days
And it’s an artist channel not a podcast
I asked claude to explain some if what this is doing and compare from yesterday to today. I am not even sure if I understand half of the tech behind it.
On the review page, the progress bar does not move along, so when I report part of a song, I cannot see the exact time where there is an issue. you think that if I can select a section of the lyrics and say that this part is not synced that it will be easier?
You also saw this link on ivelt that has booklets of many cd’s and this google drive link?
Most have the text in the pdf and done even need OCR.
Learning as I go
optimizing more and more, mostly pushing resources right now toward the Yiddish model.
Added, keep in mind that I have the user submissions quarantined, so when someone approves it’s not automatic.
Took everything from there a while ago
Found a bunch of telegram Yiddish music and lyrics channels those are being ingested and verified now
They also help the model
Still far away from being able to WRITE Yiddish lyrics, but to verify and sync so far looking like this may work
Doesn’t ivrit.ai have a yiddish speech to text? i think that the wonder phone uses that for stt.
I’ll look
Mozilla also has https://mozilladatacollective.com/datasets/cmqi72hiv00hqo5077ilqlp5y
Seems great
Happy to hear. Great as in better what you had till now?
You are only using 19 gb of RAM, will this go up?
Yes, after benchmarks I reconsidered alot, using it as base model now.
Processes use whatever they need then use ram as cache
Tested a few, what if lyrics are super duper, but nikud are off?
You can make a note, if perfect actually write PERFECT SYNC (differentiate line by line and word by word), and you can mention the nekodos too (those weren’t made by me though so we will see).
They are lashon kodesh instead of Hebrew

