Can Pyro be used to model Speech as cont. HMM model?

Basically I want to use pyro for cont. distribution. I am trying to model DNN-HMM model for phoneme sequence output. The DNN part is easy but how to model the HMM part in pyro ?? Is it possible?

any help here?

it would be much easier to answer your question if you were much more specific about the exact model structure you’re interested in. “dnn-hmm model” doesn’t give enough information to be of much use.

Looking to build a TDNN-HMM model. TDNN will give posterier probabilities to a triphone context dependent HMM . Will need to do viterbi training and viterbi decoding. TDNN will be trained using MMI/sMBR loss.

i’m afraid stringing together a few undefined acronyms isn’t helpful. something more like “we have a sequence of discrete latent random variables whose conditional probability […]. conditional on each discrete random variable […]. etc etc etc”

Hi @pd123 could you link to a paper or two describing your model?

Here are some general answers:

  • Pyro can model some combinations of DNN and HMM
  • Pyro’s HMM can have continuous-valued observation
  • Pyro’s HMM inference supports marginalized inference (sum-product), not viterbi inference (max-product). Marginalized inference is more accurate but restricts the class of DNN+HMM models that can be used.
  • Pyro’s HMM predictions currently do not provide access to the discrete latent states. We should be fixing this by the next Pyro release.