Phonon-2: most accurate open speech recognition model in a 164 MB download | Dark Hacker News