OM-LSA(Optimally-modified log-spectral amplitude)是经典的降噪算法,这里做一个学习总结,把中间的一些公式推导一下,写一些自己的想法,水平有限,欢迎指正。
LSA估计增益
OM-LSA(Optimally-modified log-spectral amplitude)从名字上就可以看出来估计的是对数幅度谱,LSA做为最小优化目标,即,
e
=
E
{
(
l
o
g
(
A
^
(
k
,
l
)
)
−
l
o
g
(
A
(
k
,
l
)
)
)
2
}
(1)
e=E\{(log(\hat{A}(k,l))- log(A(k,l)))^2\} \tag{1}
e=E{(log(A^(k,l))−log(A(k,l)))2}(1)
这里使用的其实是贝叶斯MSE准则,通过对
l
o
g
(
A
^
(
k
,
l
)
)
log(\hat{A}(k,l))
log(A^(k,l))求导得到,
l
o
g
(
A
^
(
k
,
l
)
)
=
E
(
l
o
g
(
A
(
k
,
l
)
∣
Y
(
k
,
l
)
)
)
A
^
(
k
,
l
)
=
e
x
p
(
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
)
)
(2)
log(\hat{A}(k,l))=E(log(A(k,l)|Y(k,l))) \\ \hat{A}(k,l) = exp(E(log A(k,l)|Y(k,l))) \tag{2}
log(A^(k,l))=E(log(A(k,l)∣Y(k,l)))A^(k,l)=exp(E(logA(k,l)∣Y(k,l)))(2)
每一帧可以分为存在语音和不存在语音两种情况,即,
H
0
(
k
,
l
)
:
Y
(
k
,
l
)
=
D
(
k
,
l
)
H
1
(
k
,
l
)
:
Y
(
k
,
l
)
=
X
(
k
,
l
)
+
D
(
k
,
l
)
(3)
\begin{aligned} H_0(k,l):Y(k,l) &= D(k,l) \\ H_1(k,l):Y(k,l)&=X(k,l) + D(k,l) \end{aligned} \tag{3}
H0(k,l):Y(k,l)H1(k,l):Y(k,l)=D(k,l)=X(k,l)+D(k,l)(3)
式(2)中需要求解E(log A(k,l)|Y(k,l)),在存在语音和不存在语音两种假设情况下可以写成,
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
)
=
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
1
(
k
,
l
)
)
p
(
k
,
l
)
+
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
0
(
k
,
l
)
)
(
1
−
p
(
k
,
l
)
)
(4)
\begin{aligned} E(log A(k,l)|Y(k,l)) &= E(log A(k,l)|Y(k,l),H_1(k,l))p(k,l) \\&+ E(log A(k,l)|Y(k,l),H_0(k,l))(1-p(k,l)) \end{aligned} \tag{4}
E(logA(k,l)∣Y(k,l))=E(logA(k,l)∣Y(k,l),H1(k,l))p(k,l)+E(logA(k,l)∣Y(k,l),H0(k,l))(1−p(k,l))(4)
将式(4)带人(2),并且利用
e
x
y
=
(
e
x
)
y
,
e
x
+
y
=
e
x
e
y
e^{xy}=(e^x)^y,e^{x+y}=e^xe^y
exy=(ex)y,ex+y=exey的性质,可以得到,
A
^
(
k
,
l
)
=
e
x
p
(
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
1
(
k
,
l
)
)
)
p
(
k
,
l
)
+
e
x
p
(
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
0
(
k
,
l
)
)
)
(
1
−
p
(
k
,
l
)
)
(5)
\begin{aligned} \hat{A}(k,l)&= exp(E(log A(k,l)|Y(k,l),H_1(k,l)))^{p(k,l)} \\&+ exp(E(log A(k,l)|Y(k,l),H_0(k,l)))^{(1-p(k,l))} \end{aligned} \tag{5}
A^(k,l)=exp(E(logA(k,l)∣Y(k,l),H1(k,l)))p(k,l)+exp(E(logA(k,l)∣Y(k,l),H0(k,l)))(1−p(k,l))(5)
当语音不存在时,使用一个最小值超参进行约束,即,
e
x
p
(
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
0
(
k
,
l
)
)
)
=
G
m
i
n
∣
Y
(
k
,
l
)
∣
(6)
exp(E(log A(k,l)|Y(k,l),H_0(k,l)))=G_{min}|Y(k,l)| \tag{6}
exp(E(logA(k,l)∣Y(k,l),H0(k,l)))=Gmin∣Y(k,l)∣(6)
当语音存在时,可以推导出,
e
x
p
(
E
(
l
o
g
A
(
k
,
l
)
∣
Y
(
k
,
l
)
,
H
1
(
k
,
l
)
)
)
=
G
H
1
∣
Y
(
k
,
l
)
∣
G
H
1
=
ζ
(
k
,
l
)
1
+
ζ
(
k
,
l
)
∫
v
(
k
,
l
)
∞
1
2
e
−
t
t
d
t
(7)
exp(E(log A(k,l)|Y(k,l),H_1(k,l)))=G_{H_1}|Y(k,l)| \\ G_{H_1}=\frac{\zeta(k,l)}{1+\zeta(k,l)}\int_{v(k,l)}^\infty \frac{1}{2}\frac{e^{-t}}{t}dt \tag{7}
exp(E(logA(k,l)∣Y(k,l),H1(k,l)))=GH1∣Y(k,l)∣GH1=1+ζ(k,l)ζ(k,l)∫v(k,l)∞21te−tdt(7)
其中
p
(
k
,
l
)
p(k,l)
p(k,l)为条件语音存在概率,
ζ
(
k
,
l
)
\zeta(k,l)
ζ(k,l)为先验信噪比,
γ
(
k
,
l
)
\gamma(k,l)
γ(k,l)为后验信噪比,
v
(
k
,
l
)
=
γ
(
k
,
l
)
ζ
(
k
,
l
)
1
+
ζ
(
k
,
l
)
v(k,l)=\frac{\gamma(k,l)\zeta(k,l)}{1+\zeta(k,l)}
v(k,l)=1+ζ(k,l)γ(k,l)ζ(k,l)是先验后验信噪比的函数。将式(7),(6)带入式(5)得,
A
^
(
k
,
l
)
=
{
G
H
1
(
k
,
l
)
}
p
(
k
,
l
)
G
m
i
n
1
−
p
(
k
,
l
)
∣
Y
(
k
,
l
)
∣
=
G
(
k
,
l
)
∣
Y
(
k
,
l
)
∣
(8)
\begin{aligned} \hat{A}(k,l)&= \{G_{H_1}(k,l)\}^{p(k,l)}G_{min}^{1-p(k,l)} |Y(k,l)| \\ &= G(k,l) |Y(k,l)| \end{aligned} \tag{8}
A^(k,l)={GH1(k,l)}p(k,l)Gmin1−p(k,l)∣Y(k,l)∣=G(k,l)∣Y(k,l)∣(8)
要想得到 G ( k , l ) ) G(k,l)) G(k,l)),需要估计出条件语音存在概率和先后验信噪比(间接需要估计出底噪)。
贝叶斯定理估计条件语音存在概率
假设干净语音和噪声的短时傅里叶变换系数满足复高斯分布,且不相干,则可以得到概率密度函数为,
p
(
Y
(
k
,
l
)
∣
H
0
(
k
,
l
)
)
=
1
π
λ
x
(
k
,
l
)
e
x
p
(
−
∣
Y
(
k
,
l
)
∣
2
λ
d
(
k
)
)
,
p
(
Y
(
k
,
l
)
∣
H
1
(
k
,
l
)
)
=
1
π
(
λ
x
(
k
,
l
)
+
λ
d
(
k
,
l
)
)
e
x
p
(
−
∣
Y
(
k
,
l
)
∣
2
λ
x
(
k
)
+
λ
d
(
k
)
)
(9)
\begin{aligned} p(Y(k,l)| H_0(k,l)) &= \frac{1}{\pi\lambda_x(k,l)}exp(-\frac{|Y(k,l)|^2}{\lambda_d(k)}), \\ p(Y(k,l)| H_1(k,l)) &= \frac{1}{\pi(\lambda_x(k,l)+\lambda_d(k,l))}exp(-\frac{|Y(k,l)|^2}{\lambda_x(k)+\lambda_d(k)}) \end{aligned} \tag{9}
p(Y(k,l)∣H0(k,l))p(Y(k,l)∣H1(k,l))=πλx(k,l)1exp(−λd(k)∣Y(k,l)∣2),=π(λx(k,l)+λd(k,l))1exp(−λx(k)+λd(k)∣Y(k,l)∣2)(9)
利用贝叶斯定理,
p
(
H
1
(
k
,
l
)
∣
Y
(
k
,
l
)
)
=
p
(
Y
(
k
,
l
)
∣
H
1
(
k
,
l
)
)
p
(
H
1
)
p
(
Y
(
k
,
l
)
)
=
p
(
Y
(
k
,
l
)
∣
H
1
(
k
,
l
)
)
p
(
H
1
)
p
(
Y
(
k
,
l
)
∣
H
1
(
k
,
l
)
)
p
(
H
1
)
+
p
(
Y
(
k
,
l
)
∣
H
0
(
k
,
l
)
)
p
(
H
0
)
(10)
\begin{aligned} p(H_1(k,l)|Y(k,l))&=\frac{p(Y(k,l)|H_1(k,l))p(H_1)}{p(Y(k,l))} \\ &=\frac{p(Y(k,l)|H_1(k,l))p(H_1)}{p(Y(k,l)|H_1(k,l))p(H_1) + p(Y(k,l)|H_0(k,l))p(H_0)} \end{aligned} \tag{10}
p(H1(k,l)∣Y(k,l))=p(Y(k,l))p(Y(k,l)∣H1(k,l))p(H1)=p(Y(k,l)∣H1(k,l))p(H1)+p(Y(k,l)∣H0(k,l))p(H0)p(Y(k,l)∣H1(k,l))p(H1)(10)
将式(9)带入式(10),得到,
p
(
k
,
l
)
=
{
1
+
q
(
k
,
l
)
1
−
q
(
k
,
l
)
(
1
+
ζ
(
k
,
l
)
)
e
x
p
(
−
v
(
k
,
l
)
)
}
−
1
(11)
p(k,l)=\{1+\frac{q(k,l)}{1-q(k,l)}(1+\zeta(k,l))exp(-v(k,l)) \}^{-1} \tag{11}
p(k,l)={1+1−q(k,l)q(k,l)(1+ζ(k,l))exp(−v(k,l))}−1(11)
其中 q ( k , l ) q(k,l) q(k,l)为先验语音缺失概率, ζ ( k , l ) = λ x ( k , l ) λ d ( k , l ) \zeta(k,l)=\frac{\lambda_x(k,l)}{\lambda_d(k,l)} ζ(k,l)=λd(k,l)λx(k,l)为先验信噪比, γ ( k , l ) = ∣ Y ( k , l ) ∣ 2 λ d ( k , l ) \gamma(k,l)=\frac{|Y(k,l)|^2}{\lambda_d(k,l)} γ(k,l)=λd(k,l)∣Y(k,l)∣2为后验信噪比, v ( k , l ) = γ ( k , l ) ζ ( k , l ) 1 + ζ ( k , l ) v(k,l)=\frac{\gamma(k,l)\zeta(k,l)}{1+\zeta(k,l)} v(k,l)=1+ζ(k,l)γ(k,l)ζ(k,l)是先验后验信噪比的函数(和上面的定义一样)。
根据式(11),要想得到条件语音存在概率,需要估计出先后验信噪比(需要估计底噪)和先验语音缺失概率。
DD准则估计先验信噪比
在假设噪声估计出来的情况下,后验信噪比的估计只需要利用当前帧的能量除以噪声能量就可以了,而先验信噪比的估计需要利用DD准则进行估计,即,
ζ
^
(
k
,
l
)
=
α
G
H
1
2
(
k
,
l
−
1
)
γ
(
k
,
l
−
1
)
+
(
1
−
α
)
m
a
x
{
γ
(
k
,
l
)
−
1
,
0
}
(12)
\hat{\zeta}(k,l)=\alpha G_{H_1}^2(k,l-1)\gamma(k,l-1) + (1-\alpha)max\{\gamma(k,l)-1,0\} \tag{12}
ζ^(k,l)=αGH12(k,l−1)γ(k,l−1)+(1−α)max{γ(k,l)−1,0}(12)
式(12)的第一部分是上一帧的先验信噪比估计,
G
H
1
2
(
k
,
l
−
1
)
γ
(
k
,
l
−
1
)
=
{
G
H
1
(
k
,
l
−
1
)
∣
Y
(
k
,
l
−
1
)
∣
}
2
λ
d
(
k
,
l
−
1
)
=
λ
x
(
k
,
l
−
1
)
λ
d
(
k
,
l
−
1
)
=
ζ
(
k
,
l
−
1
)
\begin{aligned} G_{H_1}^2(k,l-1)\gamma(k,l-1)&=\frac{\{G_{H_1}(k,l-1)|Y(k,l-1)|\}^2}{\lambda_d(k,l-1)}\\ &= \frac{\lambda_x(k,l-1)}{\lambda_d(k,l-1)} \\ &= \zeta(k,l-1) \end{aligned}
GH12(k,l−1)γ(k,l−1)=λd(k,l−1){GH1(k,l−1)∣Y(k,l−1)∣}2=λd(k,l−1)λx(k,l−1)=ζ(k,l−1)
式(12)的第二部分是当前帧的先验信噪比估计,
γ
(
k
,
l
)
−
1
=
∣
Y
(
k
,
l
)
∣
2
λ
d
(
k
,
l
)
−
1
=
∣
Y
(
k
,
l
)
∣
2
−
λ
d
(
k
,
l
)
λ
d
(
k
,
l
)
=
λ
x
(
k
,
l
)
λ
d
(
k
,
l
)
=
ζ
(
k
,
l
)
\gamma(k,l)-1 = \frac{|Y(k,l)|^2}{\lambda_d(k,l)} - 1 = \frac{|Y(k,l)|^2-\lambda_d(k,l)}{\lambda_d(k,l)} = \frac{\lambda_x(k,l)}{\lambda_d(k,l)} = \zeta(k,l)
γ(k,l)−1=λd(k,l)∣Y(k,l)∣2−1=λd(k,l)∣Y(k,l)∣2−λd(k,l)=λd(k,l)λx(k,l)=ζ(k,l)
将两部分做一个平滑得到最终的先验信噪比估计。
通过先验信噪比的软决策估计先验语音缺失概率
首先将估计出来的先验信噪比进行平滑,即,
ζ
(
k
,
l
)
=
β
ζ
(
k
,
l
−
1
)
+
(
1
−
β
)
ζ
^
(
k
,
l
−
1
)
(13)
\zeta(k,l)=\beta \zeta(k,l-1) + (1-\beta)\hat{\zeta}(k,l-1) \tag{13}
ζ(k,l)=βζ(k,l−1)+(1−β)ζ^(k,l−1)(13)
接着在频域进行综合,考虑频带间的影响,定义
ζ
λ
(
k
,
l
)
=
∑
i
=
−
w
λ
w
λ
h
λ
(
i
)
ζ
(
k
−
i
,
l
)
\zeta_\lambda(k,l)=\sum_{i=-w_\lambda}^{w_\lambda}h_\lambda(i)\zeta(k-i,l)
ζλ(k,l)=i=−wλ∑wλhλ(i)ζ(k−i,l)
其中
h
λ
h_\lambda
hλ为hanning窗,下标
λ
\lambda
λ可以等于"local",也可以等于"global",当
w
λ
w_\lambda
wλ=1时,为"local",代表局部信噪比的平均,当
w
λ
w_\lambda
wλ=15时,为"global",代表全局信噪比的平均,根据
ζ
λ
(
k
,
l
)
\zeta_\lambda(k,l)
ζλ(k,l)的数值和超参
ζ
m
i
n
\zeta_{min}
ζmin,
ζ
m
a
x
\zeta_{max}
ζmax定义,
p
λ
(
k
,
l
)
=
{
0
,
if
ζ
λ
(
k
,
l
)
≤
ζ
m
i
n
1
,
ζ
λ
(
k
,
l
)
≥
ζ
m
a
x
l
o
g
(
ζ
λ
(
k
,
l
)
/
ζ
m
i
n
)
l
o
g
(
ζ
m
a
x
/
ζ
m
i
n
)
,
o
t
h
e
r
w
i
s
e
(14)
p_\lambda(k,l)=\begin{cases} & 0, \text{if} \zeta_\lambda(k,l) \leq \zeta_{min} \\ & 1, \zeta_\lambda(k,l) \geq \zeta_{max}\\ & \frac{log(\zeta_\lambda(k,l)/\zeta_{min})}{log(\zeta_{max}/\zeta_{min})}, otherwise \end{cases} \tag{14}
pλ(k,l)=⎩⎪⎨⎪⎧0,ifζλ(k,l)≤ζmin1,ζλ(k,l)≥ζmaxlog(ζmax/ζmin)log(ζλ(k,l)/ζmin),otherwise(14)
为了进一步在噪声帧加强噪声抑制,定义,
ζ
f
r
a
m
e
(
l
)
=
m
e
a
n
1
≤
k
≤
M
/
2
+
1
{
ζ
(
k
,
l
)
}
\zeta_{frame}(l)=\underset{1\leq k \leq M/2+1}{mean}\{\zeta(k,l)\}
ζframe(l)=1≤k≤M/2+1mean{ζ(k,l)}
u ( l ) = { 0 , if , ζ f r a m e ( l ) ≤ ζ p e a k ζ m i n 1 , if , ζ f r a m e ( l ) ≥ ζ p e a k ζ m a x l o g ( ζ f r a m e ( l ) / ζ p e a k ( l ) / ζ m i n ( l ) ) l o g ( ζ m a x / ζ m i n ) , o t h e r w i s e (15) u(l)=\begin{cases} &0,\text{if}, \zeta_{frame}(l) \leq \zeta_{peak}\zeta_{min} \\ & 1, \text{if}, \zeta_{frame}(l) \geq \zeta_{peak}\zeta_{max} \\ & \frac{log(\zeta_{frame}(l)/\zeta_{peak}(l)/\zeta_{min}(l))}{log(\zeta_{max}/\zeta_{min})}, otherwise \end{cases} \tag{15} u(l)=⎩⎪⎨⎪⎧0,if,ζframe(l)≤ζpeakζmin1,if,ζframe(l)≥ζpeakζmaxlog(ζmax/ζmin)log(ζframe(l)/ζpeak(l)/ζmin(l)),otherwise(15)
p
f
r
a
m
e
p_{frame}
pframe的计算如下图所示,
先验语音缺失概率根据下式求得,
q
^
(
k
,
l
)
=
1
−
p
l
o
c
a
l
(
k
,
l
)
p
g
l
o
b
a
l
(
k
,
l
)
p
f
r
a
m
e
(
l
)
(16)
\hat{q}(k,l)= 1 - p_{local}(k,l)p_{global}(k,l)p_{frame}(l) \tag{16}
q^(k,l)=1−plocal(k,l)pglobal(k,l)pframe(l)(16)
MCRA方法估计底噪
常见的噪声估计方法有递归平滑,最小值跟踪,直方图统计和分位数噪声估计等方法,这里的MCRA(Minima controlled recursive averaging)算法使用了递归平均和最小值跟踪相结合的方法,这里的最小值跟踪体现在语音存在概率是由最小值跟踪确定的。
递归平滑
在噪声段进行语音平滑处理,在语音段不更新噪声,则,
H
0
′
(
k
,
l
)
:
λ
d
^
(
k
,
l
+
1
)
=
α
d
λ
d
^
(
k
,
l
)
+
(
1
−
α
d
)
∣
Y
(
k
,
l
)
∣
2
H
1
′
(
k
,
l
)
:
λ
d
^
(
k
,
l
+
1
)
=
λ
d
^
(
k
,
l
)
(17)
\begin{aligned} H_0^{'}(k,l):\hat{\lambda_d}(k,l+1) &= \alpha_d\hat{\lambda_d}(k,l) + (1-\alpha_d)|Y(k,l)|^2 \\ H_1^{'}(k,l):\hat{\lambda_d}(k,l+1) &= \hat{\lambda_d}(k,l) \end{aligned} \tag{17}
H0′(k,l):λd^(k,l+1)H1′(k,l):λd^(k,l+1)=αdλd^(k,l)+(1−αd)∣Y(k,l)∣2=λd^(k,l)(17)
α
d
\alpha_d
αd是平滑参数,数值过大会导致跟踪较慢,数值过小,容易产生音乐噪声。利用条件语音存在概率
p
′
(
k
,
l
)
p^{'}(k,l)
p′(k,l)进行综合,可得,
λ
d
^
(
k
,
l
+
1
)
=
(
α
d
λ
d
^
(
k
,
l
)
+
(
1
−
α
d
)
∣
Y
(
k
,
l
)
∣
2
)
(
1
−
p
′
(
k
,
l
)
)
+
λ
d
^
(
k
,
l
)
p
′
(
k
,
l
)
=
α
~
d
λ
d
^
(
k
,
l
)
+
(
1
−
α
~
d
)
∣
Y
(
k
,
l
)
∣
2
(18)
\begin{aligned} \hat{\lambda_d}(k,l+1) &= (\alpha_d\hat{\lambda_d}(k,l) + (1-\alpha_d)|Y(k,l)|^2)(1-p^{'}(k,l)) + \hat{\lambda_d}(k,l)p^{'}(k,l) \\ & = \tilde\alpha_d\hat{\lambda_d}(k,l) + (1-\tilde\alpha_d)|Y(k,l)|^2 \end{aligned} \tag{18}
λd^(k,l+1)=(αdλd^(k,l)+(1−αd)∣Y(k,l)∣2)(1−p′(k,l))+λd^(k,l)p′(k,l)=α~dλd^(k,l)+(1−α~d)∣Y(k,l)∣2(18)
其中, α d ~ = α d + ( 1 − α d ) p ′ ( k , l ) \tilde{\alpha_d}=\alpha_d + (1-\alpha_d)p^{'}(k,l) αd~=αd+(1−αd)p′(k,l)
最小值控制语音存在概率
首先对幅度谱在频域进行平滑,得,
S
f
(
k
,
l
)
=
∑
i
=
−
w
w
b
(
i
)
∣
Y
(
k
−
i
,
l
)
∣
2
(19)
S_f(k,l)=\sum_{i=-w}^{w}b(i)|Y(k-i,l)|^2 \tag{19}
Sf(k,l)=i=−w∑wb(i)∣Y(k−i,l)∣2(19)
接着在时域进行平滑,
S
(
k
,
l
)
=
α
s
S
(
k
,
l
−
1
)
+
(
1
−
α
s
)
S
f
(
k
,
l
)
(20)
S(k,l)=\alpha_s S(k,l-1) + (1-\alpha_s )S_f(k,l) \tag{20}
S(k,l)=αsS(k,l−1)+(1−αs)Sf(k,l)(20)
利用最小值跟踪法求出
S
m
i
n
(
k
,
l
)
S_{min}(k,l)
Smin(k,l),并且定义比值
S
r
(
k
,
l
)
=
S
(
k
,
l
)
/
S
m
i
n
(
k
,
l
)
S_r(k,l)=S(k,l)/S_{min}(k,l)
Sr(k,l)=S(k,l)/Smin(k,l),如果
S
r
(
k
,
l
)
S_r(k,l)
Sr(k,l)大于阈值,则令
I
(
k
,
l
)
=
1
I(k,l)=1
I(k,l)=1,否则为0,条件语音存在概率根据I(k,l)平滑得到,即,
p
′
(
k
,
l
)
=
α
p
p
′
(
k
,
l
−
1
)
+
(
1
−
α
)
I
(
k
,
l
)
(21)
p^{'}(k,l)=\alpha_pp^{'}(k,l-1) + (1-\alpha_)I(k,l) \tag{21}
p′(k,l)=αpp′(k,l−1)+(1−α)I(k,l)(21)
将式(21)带入式(18)得到最终的噪声估计值。
总结
算法流程为:首先使用MCRA方法估计出噪声,接着估计出后验信噪比和先验信噪比(DD准则),利用先验信噪比的软决策估计出先验语音缺失概率,之后利用贝叶斯准则求出条件语音存在概率,将前面求出的值带入增益函数表达式得到增益值。