first, word2vec vectors =!= transformer style vectors. The first is per word, zero context. The latter is mostly about context, for all the words in the context.
second, word2vec embeddings can be replaced instead of the word, straight drop-in. Works very well. Even for single word as an input. Works very badly if context is the key thing. So {sushi, vomit} is {[300 floats], [300 floats]} where the first 300 floats do not know they are associated to ugliness.
TODO Move around overlays in StreamElements so that the text doesn’t overlap
TODO Look at your own vods to see how the overlays/stream stuff can get fixed for chat
TODO See if bitrate is hampering video quality on stream