Re: [SoX-users] How can I get SNR value of an audio?

[David G. <gr...@ld...>]

xl...@so... said:
> Is there some methods in SoX that I can use to get the SNR value of an
> audio? If not, could you give me some advice to do this? 

ha...@st... said:
> First you need to specify what part of your audio constitute the
> "signal" and what constitutes "noise". 

What Jan Stary said is correct, but also: if your audio file is
essentially "clean" (all portions are either "quiet" or "good signal",
with no portions that have loud noises, like doors slamming, tapping on
the microphone, etc), then the "stats" effect will report the following:

 command line:

  sox some_audio.file -n stats

 output (for a sample audio file I happened to have available):

               Overall     Left      Right
  DC offset  -0.000001 -0.000001 -0.000001
  Min level  -0.980347 -0.980347 -0.605347
  Max level   0.980347  0.980347  0.542847
  Pk lev dB      -0.17     -0.17     -4.36
  RMS lev dB    -24.54    -21.79    -33.83
  RMS Pk dB      -7.06     -7.06    -16.96
  RMS Tr dB     -88.95    -72.85    -88.95
  Crest factor       -     12.05     29.76
  Flat factor     1.09      1.10      0.00
  Pk count          60       118         2
  Bit-depth      14/14     14/14     14/14
  Num samples    4.84M
  Length s     605.040
  Scale max   1.000000
  Window s       0.050

In this example, the default window size of 50 msec ("Window s  0.050") is
used to compute the RMS dB for the loudest window ("RMS Pk dB -7.06") and
for the quietest ("trough") window ("RMS Tr dB -88.95"), and if my own
understanding is correct, the dB SNR can be estimated as the difference
between those two values (in this case, 81.89 overall, 65.79 on the left 
channel, 71.99 on the right channel).

But this would be a "best case" type of estimate -- that is, it tells 
you what the maximum possible SNR could be for the file.  The actual SNR 
is likely to be lower, if you take the average RMS dB from a bunch of 
"quiet" portions and compare that to the average RMS dB from a bunch of 
"typical signal" portions.

Check the sox documentation regarding the "stats" effect for more info.

	Dave Graff