（一）网络结构

input-> C1 –>softmax –>S2 –> C3 –>softmax –>S4 –>FC5 –> softmax –> FC6

（二）数据预处理

1.归一化：提取音频.wav的对数频谱值作为特征值（d=129），并归一化之；

2.扩帧：对每一帧左右各扩5帧，扩帧后的维度 D =d*(5*2+1) =d*11；

（三）初始化各网络层

1. conv1

① in_dim = D

② stride1 = d

③ num_patch1 = 1+（stride1 - patch1_dim）/patch1_step

④ out_dim = num_filters1 *num_patch1

2. mpool1

① in_dim = out_dim (1.④)

② num_pool1 = num_patch1 / pool1_size

③ out_dim = num_filters1 *num_pool1

3.conv2

① in_dim = out_dim (2.③)

② stride2 = num_filters1 *num_pool1

③ patch2_dim = patch2_dim *num_filters1

④ patch2_step = num_filters1

⑤ num_patch2 = 1+（stride2 – patch2_dim）/patch2_step

⑥ out_dim = num_filters2 *num_patch2

4.mpool2

① in_dim = out_dim (3.⑥)

② num_pool2 = num_patch2 / pool2_size

③ out_dim = num_filters2 *num_pool2

5.FC1 & FC2

in_dim = out_dim = d

Kaldi + CNN for Speech Enhancement