Per https://blog.tensorflow.org/2020/03/introducing-webassembly-backend-for-tensorflow-js.html and https://blog.tensorflow.org/2020/09/supercharging-tensorflowjs-webassembly.html it sounds like merely switching the tensorflow backend to 'wasm' could result in a nice speedup.
I've tried updating to a more recent TF.js but didn't figure out the invocations to run from WASM yet. Will share if I get it running.