mutiann

'''
Implementation for https://arxiv.org/abs/1906.00672

Tips: The code could be directly used in place of BadahnauMonotonicAttention in Tensorflow codes. Similar to its 
base class in the Tensorflow seq2seq codebase,  you may use "hard" for hard inference, or "parallel" for training or 
soft inference. "recurrent" mode in BadahnauMonotonicAttention is not supported. 

If you have already trained another model using BadahnauMonotonicAttention, the model could be reused, otherwise you 
possibly have to tune the score_bias_init, which, similar to that in Raffel et al., 2017, is determined a priori to 
suit the moving speed of the alignments, i.e. speed of speech of your training corpus in TTS cases. So