Unfortunately, single digit millisecond is quite slow for n-gram language models. Weighted Finite State Transducers are where they really shine; you tend to wa

Unfortunately, single digit millisecond is quite slow for n-gram language models. Weighted Finite State Transducers are where they really shine; you tend to want to do massive parallel operations, and you want to do the without being bottle necked on memory access. I think these challenges make this framework challenging to adopt.