ref: d96b541d0c4a9bfe42413a99e07691e249c2fc42
dir: /test/README/
Here are some helper programs for checking filter and rate conversion
accuracy of sox. The programs in the test subdirectory which I
(Stanley J. Brooks) wrote are covered by the GPL. See the
../LICENSE.GPL file for details. Sox is also covered by the GPL.
To use this stuff, first cd to this test subdirectory and run make,
which should produce the corr, model, and lmodel executables.
Now to test something...
Say you want to compare the 'band' effect to the 'filter' effect...
./ltest.pl band 800 200 >A
./ltest.pl filter 600-1000 >B
gnuplot responseAB
plots the response curves with y being power-gain in dB,
and x the frequency.
For another, say to compare response and error of resample
rate conversion from 8000 ->22050 samples per second,
with linear (default) interpolation versus -qs quadratic
interpolation with Nuttall window...
./ltest.pl -l resample >A
./ltest.pl -l resample -qs 0.80 0 >B
gnuplot plotAB
The script wtest.pl is for testing sox's adpcm and gsm wav support.
It accepts one of these options:
-t Use 'toast' gsm for compress/decompress
-a Use sox (MS) ADPCM
-i Use sox IMA ADPCM
-g Use sox wav-gsm6.10
If using sox, you may also append an optional effect, eg:
./wtest -g filter 0-3500 >A
will apply a lowpass filter with 6dB corner at 3500 Hz before the
compression phase.
Then, as above, 'gnuplot plotA' will graph dB gain and dB error-level
vs. freq for you.
-------------------------------------------------------------------
About ltest:
-l means use 32-bit signed samples, otherwise 16-bit signed is used.
The other parameters are fed into sox as the 'effect' with parameters.
The rates 8000:22050 are in the script, but you can edit a perl script,
I hope. just change the ($rate0,$rate1)=(8000,22050) line near the top.
What ltest does:
It uses sox to synthesize a sineusoid input file for the
frequencies 0.01 ... 0.99 of the Nyquist frequency. This input
has:
400 samples samples of silence,
4000 samples with smooth envelope rising to volume -v0.5
16000 samples at -v0.5
4000 samples with smooth envelope falling to 0,
400 sample more of silence.
The rising/falling envelopes are shaped like rising, falling portions
of the (1-cos(x)) function.
Then the filter or rate-change effect is applied to this i0.xx.xx file
to give an output file.
The output file is examined by the 'model' or 'lmodel' program to analyse
response level and error level.
model works as follows:
step 1: the entire sample file is read in, and the center-of-gravity of
the squared samples is found. This is the time offset which should
correspond to the center of the filtered/resampled tone-pulse.
step 2: let N = 16000*(rate1/rate0) be the number of samples at output rate
which would correspond to the 16000 sample duration at max volume.
We focus on the samples between
(center - 0.3*N) and (center + 0.9*N)
where the transient effects of attack/release envelope should be
small.
We do a least-squares fit of a sinusoid at the adjusted frequency
to this segment of 0.6*N samples, and print out the component
s2max which can be explained, and also the rms level of the 'error'
or unexplained part.
That's pretty much it... the ltest perl-script glues it together and outputs
adjusted data which the gnuplot will like.
Modify the perl and gnuplot scripts to suit your needs.