I understand that the combination of MFCCs, deltas and double deltas is a good feature to be used with HMMs for keyword detection problems. HMMs are limited by Markov Property and this limitation is overcome by using deltas and double-deltas.
But in the case of LSTMs, they can model long term dependencies very easily. So, doesn't that make the delta and double delta features redundant while using LSTM as the model?