On-device CPU inference is the real flex here. Optimization probably mattered as much as modeling.
My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
41–50 of 57 posts
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#42[dead]
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#43This feels adjacent to voice conversion research, but with stricter latency constraints.
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#44The lack of parallel accent data makes this fundamentally unsupervised. Curious if this leans more on latent disentanglement than direct supervision.
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#45[dead]
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#46[dead]
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#47Without full utterance context, homophones must be tricky. How do you avoid semantic drift?
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#48Edge-side CPU inference is the quiet power move. Feels like the engineering grind on optimization carried just as much weight as the model architecture itself.
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#49[dead]
Re: My accent costs me 30 IQ points on Zoom. So we built an ML model to fix it
#50[dead]