SNR計算
S1 = 原始乾淨語音
N1 = 噪聲.
S2 = S1 + N1 (帶噪語音)
S3 = 增強後的語音(使用某種算法降噪)
N2 = S3 - S1 (增強後語音中的殘留噪聲)
SNR的計算公式爲(dB)
S N R = 10 log 10 ∥ s i g n a l ∥ 2 ∥ n o i s e ∥ 2
SNR = 10{\log _{10}}\frac{{{{\left\| {signal} \right\|}^2}}}{{{{\left\| {noise} \right\|}^2}}}
S N R = 1 0 log 1 0 ∥ n o i s e ∥ 2 ∥ s i g n a l ∥ 2
SDR的計算公式爲
S D R = 10 log 10 ∥ X c ∥ 2 ∥ X − X c ∥ 2
SDR = 10{\log _{10}}\frac{{{{\left\| {Xc} \right\|}^2}}}{{{{\left\| {X - Xc} \right\|}^2}}}
S D R = 1 0 log 1 0 ∥ X − X c ∥ 2 ∥ X c ∥ 2
其中X c X_c X c 爲帶噪語音中的乾淨分量,X X X 爲帶噪語音,X c − X X_c - X X c − X 爲帶噪語音中的噪聲分量。計算SNR提升量:
S N R ( a f t e r E n h a n c e d ) − S N R ( b e f o r e E n h a n c e d ) = 10 log 10 ∥ S 1 ∥ 2 ∥ N 2 ∥ 2 − 10 log 10 ∥ S 1 ∥ 2 ∥ N 1 ∥ 2
SNR(afterEnhanced) - SNR(beforeEnhanced) = 10{\log _{10}}\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {N2} \right\|}^2}}} - 10\log 10\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {N1} \right\|}^2}}}
S N R ( a f t e r E n h a n c e d ) − S N R ( b e f o r e E n h a n c e d ) = 1 0 log 1 0 ∥ N 2 ∥ 2 ∥ S 1 ∥ 2 − 1 0 log 1 0 ∥ N 1 ∥ 2 ∥ S 1 ∥ 2
S D R ( a f t e r E n h a n c e d ) − S D R ( b e f o r e E n h a n c e d ) = 10 log 10 ∥ S 1 ∥ 2 ∥ S 3 − S 1 ∥ 2 − 10 log 10 ∥ S 1 ∥ 2 ∥ S 2 − S 1 ∥ 2 = 10 log 10 ∥ S 1 ∥ 2 ∥ N 2 ∥ 2 − 10 log 10 ∥ S 1 ∥ 2 ∥ N 1 ∥ 2
\begin{array}{l}
SDR(afterEnhanced) - SDR(beforeEnhanced) = 10{\log _{10}}\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {S3 - S1} \right\|}^2}}} - 10\log 10\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {S2 - S1} \right\|}^2}}}\\
= 10{\log _{10}}\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {N2} \right\|}^2}}} - 10\log 10\frac{{{{\left\| {S1} \right\|}^2}}}{{{{\left\| {N1} \right\|}^2}}}
\end{array}
S D R ( a f t e r E n h a n c e d ) − S D R ( b e f o r e E n h a n c e d ) = 1 0 log 1 0 ∥ S 3 − S 1 ∥ 2 ∥ S 1 ∥ 2 − 1 0 log 1 0 ∥ S 2 − S 1 ∥ 2 ∥ S 1 ∥ 2 = 1 0 log 1 0 ∥ N 2 ∥ 2 ∥ S 1 ∥ 2 − 1 0 log 1 0 ∥ N 1 ∥ 2 ∥ S 1 ∥ 2
從公式上看,兩者完全相同。
SDR是(輸入信號的功率)和(輸入信號與增強信號之差的功率比),與SNR是一樣的,在語音增強中,他們都反應了整體的性能。SDR的性能可以反應SNR的性能.
另外,在下面文獻中也有類似的結論
Huang, Po-Sen, et al. “Joint optimization of masks and deep recurrent neural networks for monaural source separation.” IEEE/ACM Transactions on Audio, Speech, and Language Processing (TASLP) 23.12 (2015): 2136-2147.
SDR是輸入信號的功率與輸入信號與重構信號之差的功率之比。因此,SDR與經典的測量“信噪比”(SNR)完全相同,SDR反映了整體的分離性能。