“Imprecise” language models are smaller, speedier, and nearly as accurate
“Imprecise” language models are smaller, speedier, and nearly as accurate
“Imprecise” language models are smaller, speedier—and nearly as accurate
I maintain that prior to Siri, the Dragon-based iOS Voice Command feature (on iPhone 4) was accurate and functional for most purposes, and did on-device parsing. If I asked “Call Mums Mobile” or “Play Jonathan Coulton” it would parse my contacts for a contact named (or nicknamed) Mum and parse my music library for an artist who’s name matched “ˈdʒɑnəθən ˈkoʊltn̩”.
For whatever reason, Siri records my voice, uploads it to the cloud and decides that I wanted to call “Moms homestyle kitchen” and play Drake.
Makes me wonder how much correlation there is between “good enough” hardware emulation (e.g. old consoles) and “good enough” language models in terms of effort and computational demands