Comment by usef-

I think they're claiming it's achieved by text models, not voice models, fwiw.