Is this something worth trying for speech ?

[ghost]

As a research work, I am trying to build a DNN-HMM model for low resource speech. Do you think I can use this for the hmm implementation for training and decoding since there is no library in tensorflow supporting HMM?

I am really thinking of scaling this hybrid model and keeping it neater than a traditional kaldi pipeline though accuracy is something I am not worried about currently.

Please lemme know.

[dwiel]

I dont see why not. If you want to explore this, I recommend benchmarking it for your particular use case. I have found it can become slow as the number of states increases.