I referred to the TDNN, TDNN-LSTM, TDNN-Attention models provided by Kaldi. I wanted to use this to implement the model with Pytorch, but it was difficult to implement the following:
- delay : the delay to be used in the recurrence of LSTMs
- decay-time : an approximate maximum on how many frames can be remembered via summation into the cell contents
- projection-dim : the output dimension of the recurrent-projection-matrix
Has anyone ever had a github repository for reference or implemented such a model? I’d appreciate it if you could show me the implementation code.