Fast/Approximate Sampling for HMMs/HDP-HMMs

I’d say look around for an HMM package to actually compute a likelihood and run Turing on top of that.