Algorithm 2: Processing DWT output signals through CNN-BLSTM Algorithm

Input: speech signals, Deep learning parameter (batch size, feature dimension, classes, train test ratio).

Output: enhanced speech signal with recognition rate performance.

Step 1: capture speech signals by using DMA microphone array

Step 2: Apply an analogue to digital converter to convert an analogue signal into a digital signal.

Step 3: apply wavelet transform by applying $\mathit{X}\left(\mathit{a},\mathit{b}\right)=\frac{1}{\sqrt{\mathit{a}}}{\int }_{-\infty }^{\infty }\psi \left(\frac{\mathit{t}-\mathit{b}}{\mathit{b}}\right)\mathit{x}\left(\mathit{t}\right)\mathit{dt}$• Decompose signal into LL, HL, LH, and HH bands by computing the wavelet coefficients as ${\mathit{c}}_{\mathit{jk}}=\left[{\mathit{W}}_{\psi }\mathit{f}\right]\left(2^{-\mathit{j}},\mathit{k}{2}^{-\mathit{j}}\right)$

Step 4: Input these coefficients to deep learning

• Process through convolutional layers ${\mathit{n}}_{\mathit{out}}=\left[\frac{{\mathit{n}}_{\mathit{in}}+2\mathit{p}-\mathit{k}}{\mathit{s}}\right]+1$, nin denotes the input attributes, nout denotes the output features, k convolution kernel size, p padding size, s is the stride

• Process the convolved data through pooling layer ${\mathit{h}}_{\mathit{xy}}^{\mathit{l}}={\mathbf{max}}_{\mathit{i}=0,\dots \mathit{s},\mathit{j}=0,..\mathit{s}}{\mathit{h}}_{\left(\mathit{x}+1\right)\left(\mathit{y}+\mathit{j}\right)}^{\mathit{l}-1}$

• Perform linearization by applying linear layer

• Apply BiLSTM layer

• Process the memory unit data through fully connected layer zl = Wlhl−1

• Soft max layer $\mathit{softmax}\left({\mathit{z}}_{\mathit{i}}\right)=\frac{{\mathit{e}}^{{\mathit{z}}_{\mathit{i}}}}{{\sum }_{\mathit{j}}{\mathit{e}}^{{\mathit{z}}_{\mathit{j}}}}$

Step 5: obtain the final output speech data and measure the performance