首先,k时刻w频点的幅值由语音信号和噪声信号组成,我们的目的是减少噪声那部分的幅值。假设k时刻w频点语音、噪声幅值比例为 η k ( w ) ∈ [ 0 , 1 ] \eta_k(w)\in[0,1] ηk(w)∈[0,1],干净语音信号为 S k ( w ) S_k(w) Sk(w),噪声信号 N k ( w ) N_k(w) Nk(w),那么,干净语音信号的功率谱为: ∑ w ∑ k ∣ S k ( w ) ∣ 2 \sum_w \sum_k|S_k(w)|^2 ∑w∑k∣Sk(w)∣2,语音片段的噪声为 ∣ S k ( w ) ∣ ( 1 − η k ( w ) ) |S_k(w)|(1-\eta_k(w)) ∣Sk(w)∣(1−ηk(w)),非语音片段的噪声为 ∣ N k ( w ) ∣ η k ( w ) |N_k(w)|\eta_k(w) ∣Nk(w)∣ηk(w),于是,噪声信号的功率谱为: ∑ w ∑ k ( ∣ N k ( w ) ∣ η k ( w ) + ∣ S k ( w ) ∣ ( 1 − η k ( w ) ) ) 2 \sum_w \sum_k(|N_k(w)|\eta_k(w)+|S_k(w)|(1-\eta_k(w)))^2 ∑w∑k(∣Nk(w)∣ηk(w)+∣Sk(w)∣(1−ηk(w)))2,假设噪声信号和语音信号不相关,噪声信号的功率谱可以化简为: ∑ w ∑ k ∣ N k ( w ) ∣ 2 η k 2 ( w ) + ∣ S k ( w ) ∣ 2 ( 1 − η k ( w ) ) 2 \sum_w \sum_k|N_k(w)|^2\eta_k^2(w)+|S_k(w)|^2(1-\eta_k(w))^2 ∑w∑k∣Nk(w)∣2ηk2(w)+∣Sk(w)∣2(1−ηk(w))2
因此,整体的信噪比R可以定义为:
R
=
∑
w
∑
k
∣
S
k
(
w
)
∣
2
∑
w
∑
k
∣
N
k
(
w
)
∣
2
η
k
2
(
w
)
+
∣
S
k
(
w
)
∣
2
(
1
−
η
k
(
w
)
)
2
(D-1)
R=\frac{\sum_w \sum_k|S_k(w)|^2}{\sum_w \sum_k|N_k(w)|^2\eta_k^2(w)+|S_k(w)|^2(1-\eta_k(w))^2}\tag{D-1}
R=∑w∑k∣Nk(w)∣2ηk2(w)+∣Sk(w)∣2(1−ηk(w))2∑w∑k∣Sk(w)∣2(D-1)
所以现在目标很明确,就是要让整体的信噪比R最大。此时将(D-1)对
η
k
(
w
)
\eta_k(w)
ηk(w)求导并等于0,求得信噪比最大时的
η
k
(
w
)
\eta_k(w)
ηk(w)值:
由求导公式:
(
u
v
)
′
=
u
′
v
−
u
v
′
v
2
(\frac{u}{v})^{'}=\frac{u^{'}v-uv^{'}}{v^2}
(vu)′=v2u′v−uv′
得:
R
′
=
−
(
∑
w
∑
k
∣
S
k
(
w
)
∣
2
)
(
2
∣
N
k
(
w
)
∣
2
η
k
(
w
)
−
2
∣
S
k
(
w
)
∣
2
(
1
−
η
k
(
w
)
)
)
(
∑
w
∑
k
∣
N
k
(
w
)
∣
2
η
k
2
(
w
)
+
∣
S
k
(
w
)
∣
2
(
1
−
η
k
(
w
)
)
2
)
2
=
0
⇓
−
(
∑
w
∑
k
∣
S
k
(
w
)
∣
2
)
(
2
∣
N
k
(
w
)
∣
2
η
k
(
w
)
−
2
∣
S
k
(
w
)
∣
2
(
1
−
η
k
(
w
)
)
)
=
0
⇓
∣
N
k
(
w
)
∣
2
η
k
(
w
)
=
∣
S
k
(
w
)
∣
2
(
1
−
η
k
(
w
)
)
⇓
η
k
(
w
)
=
∣
S
k
(
w
)
∣
2
∣
N
k
(
s
)
∣
2
+
∣
S
k
(
w
)
∣
2
(D-2)
\begin{aligned}R^{'}=\frac{-(\sum_w \sum_k|S_k(w)|^2)(2|N_k(w)|^2\eta_k(w)-2|S_k(w)|^2(1-\eta_k(w)))}{(\sum_w \sum_k|N_k(w)|^2\eta_k^2(w)+|S_k(w)|^2(1-\eta_k(w))^2)^2}&=0\\ &\Downarrow\\ -(\sum_w \sum_k|S_k(w)|^2)(2|N_k(w)|^2\eta_k(w)-2|S_k(w)|^2(1-\eta_k(w)))&=0\\ &\Downarrow\\ |N_k(w)|^2\eta_k(w)&=|S_k(w)|^2(1-\eta_k(w))\\ &\Downarrow\\ \eta_k(w)&=\frac{|S_k(w)|^2}{|N_k(s)|^2+|S_k(w)|^2} \end{aligned}\tag{D-2}
R′=(∑w∑k∣Nk(w)∣2ηk2(w)+∣Sk(w)∣2(1−ηk(w))2)2−(∑w∑k∣Sk(w)∣2)(2∣Nk(w)∣2ηk(w)−2∣Sk(w)∣2(1−ηk(w)))−(w∑k∑∣Sk(w)∣2)(2∣Nk(w)∣2ηk(w)−2∣Sk(w)∣2(1−ηk(w)))∣Nk(w)∣2ηk(w)ηk(w)=0⇓=0⇓=∣Sk(w)∣2(1−ηk(w))⇓=∣Nk(s)∣2+∣Sk(w)∣2∣Sk(w)∣2(D-2)
令先验信噪比
R
k
(
w
)
=
∣
S
k
(
w
)
∣
2
∣
N
k
(
w
)
∣
2
R_k(w)=\frac{|S_k(w)|^2}{|N_k(w)|^2}
Rk(w)=∣Nk(w)∣2∣Sk(w)∣2,则(D-2)可以化简为:
η
k
(
w
)
=
R
k
(
w
)
1
+
R
k
(
w
)
(D-3)
\eta_k(w)=\frac{R_k(w)}{1+R_k(w)}\tag{D-3}
ηk(w)=1+Rk(w)Rk(w)(D-3)
从(D-3)可以定性分析出,维纳滤波器会对频谱上信噪比高的部分保持较大的权重,同时对信噪比低的部分进行抑制。如下图所示