someone got a speech model running on an apple watch.

not a toy demo. granite 4.0 1B speech just ranked FIRST on the OpenASR leaderboard.
heres whats wild about it:
• 1B parameters - half the size of granite 3.3 2B
• higher english transcription accuracy than the bigger model
• speculative decoding for faster inference on tiny hardware
• 6 languages - english, french, german, spanish, portuguese, japanese
• keyword list biasing so it actually gets names and acronyms right
the part nobody is talking about:
youre paying for whisper API calls every month while a model half the size of its predecessor is outperforming it on a device strapped to your wrist.
thats not a minor optimization. thats the entire cost structure of edge speech apps collapsing.
smaller model. better accuracy. ZERO cloud dependency.