I think speech post-training’s real bottleneck isnt compute or architecture. its that we cant label “this sounded good” at scale reliably.
MOS needs humans and that is a problem in itself but it also doesnt seem to survive contact with expressive speech because raters disagree with each other on emotional delivery more than they disagree on intelligibility . emotion is just not that objective.
And so people reach for UTMOS as the automatic proxy but UTMOS learned naturalness from mostly neutral read speech. and in my personal experience when u push a model toward expressive delivery then the score falls apart and u get penalized for succeeding. which means the reward model is fighting the objective on exactly the samples u care about.
Text got out of this trap with preference models good enough to train on. speech doesnt have that yet or maybe it does lol
@animeshingh_ west throws a biscuit and indian dogs who loves western ass latches on the biscuit . they be abusing their own father just because biden told them to
The British empire caused 165 million deaths in India in 40 years:
A new study found that British colonialism caused approximately 165 million deaths in India from 1880 to 1920, while stealing trillions of dollars of wealth
https://t.co/X6Ri5aH6lM