Advances, Systems and Applications
From: A novel privacy-preserving speech recognition framework using bidirectional LSTM
Symbol | Interpretation | Symbol | Interpretation |
---|---|---|---|
→ | The calculation process of F-LSTM | Wkh | Weight matrix of ht−1, where \(k \in \left \{f, i, \tilde {c}, o\right \}\) |
← | The calculation process of B-LSTM | Wkx | Weight matrix of xt, where \(k \in \left \{f, i, \tilde {c}, o\right \}\) |
′ | Parameter running on the edge server S1 | Bk | Bias, where \(k \in \{f, i, \tilde {c}, o\}\) |
′′ | Parameter running on the edge server S1 | uk | uk=Wk·[ht−1,xt] + Bk, where \(k \in \left \{f, i, \tilde {c}, o\right \}\) |
t | Time node during speech recognition process. | ct | the neural cell state at time t |
k | \(k \in \left \{f, i, \tilde {c}, o\right \}\) | ∇Wt | The computational gradient of weight matrix |
δk,t | The calculation error of backward propagation at time t | ∇Wkh | The computational gradient of weight matrix of ht |
ht | The output of LSTM at time t | ∇Wkx | The computational gradient of weight matrix of xt |
Wk | Weight matrix | ∇Bk | The computational gradient of bias in back propagation |