[a-dub]

Isn't there some value in being able to bench accoustic models in isolation, no matter how weak they may be, without downstream language models?