All-

Current SLMs are not capable of correcting "sound alike" errors prevalent in real-time speech recognition running on small form-factor devices (e.g. Kaldi, Whisper). Such errors become more frequent with noise and multiple talkers in the background. For robotics applications this is important, as precise, non-false-positive commands must be formulated to instruct factory robots, robotaxis, and other equipment, especially in urgent / safety-related situations.

I posted about this here:

  https://www.linkedin.com/posts/jeff-brower-1a51565...

Does anyone know of an SLM (under 16 GB mem, 250 msec token rate) that can correct the example sentence in the post ? Thanks !

-Jeff