Interesting. Can you explain how this is superior and/or different from traditional DSP filters or other non-tokenization tricks in the signal processing field?
My intuition is the following: transformers work really well for text, so we could try turning a time series into a "story" (limited vocabulary) and see what happens.