There is something wrong at line 169 in audiolib.py

[ZihaoCui]

The segmental_snr_mixer function in audiolib.py is wrong.
Because the normalization is applied, the noisescalar calculation is
noisescalar = 1 / (10**(snr/20))
not be
noisescalar = rmsclean / (10**(snr/20)) / (rmsnoise+EPS)

Based on the test file
import numpy as np
import audiolib
params = {'cfg':1, 'target_level_lower':-35,'target_level_upper':-25}
temp = [0.1,0.2,0.3,0.4,0.5]
for abs_a in range(len(temp)):
for abs_b in range(len(temp)):
base_a = np.random.binomial(1,temp[abs_a],size=(5000,1))
base_b = np.random.binomial(1,temp[abs_b],size=(5000,1))
print(temp[abs_a], temp[abs_b], audiolib.segmental_snr_mixer(params,base_a,base_b,snr=5))

and add the following code after line 169 in audiolib.py
rmsclean1, rmsnoise2 = active_rms(clean=clean, noise=noisenewlevel)
snr_test = 20*np.log10(rmsclean1/rmsnoise2)
return snr_test, rmsclean, rmsnoise

we have
0.1 0.1 (4.809261190526833, 0.3199999999999999, 0.31304951684997046)
0.1 0.2 (8.398437767900433, 0.3085449724108302, 0.45628938186199325)
0.1 0.3 (9.541254719329668, 0.32557641192199405, 0.5491812087098392)
0.1 0.4 (11.254207673313344, 0.30822070014844877, 0.6332456079595024)
0.1 0.5 (11.975890986196333, 0.3174901573277508, 0.7088018058667739)
0.2 0.1 (1.8278324324274429, 0.4436214602563766, 0.307896086366813)
0.2 0.2 (5.030355204691368, 0.44676615807377346, 0.44833023542919775)
0.2 0.3 (6.760912590556815, 0.4498888751680796, 0.5509990925582363)
0.2 0.4 (8.027724031442382, 0.4460941604639091, 0.6321392251711642)
0.2 0.5 (8.922584416928597, 0.45409250158970904, 0.7133021800050802)
0.3 0.1 (0.009673346077310881, 0.5594640292279744, 0.3149603149604724)
0.3 0.2 (3.2262700519281093, 0.5526300751859239, 0.4505552130427523)
0.3 0.3 (5.140245659899135, 0.5464430436925699, 0.5553377350765928)
0.3 0.4 (6.272472841406196, 0.5479051012721089, 0.6343500610861481)
0.3 0.5 (7.247686760565433, 0.5499090833947007, 0.7123201527403249)
0.4 0.1 (-1.350831784195107, 0.6315061361538776, 0.3039736830714132)
0.4 0.2 (2.0700880682248917, 0.6378087487640788, 0.4551922670696416)
0.4 0.3 (3.875478643051899, 0.6288083968904994, 0.5524490926773252)
0.4 0.4 (4.854129294663155, 0.6463745044476923, 0.6356099432828279)
0.4 0.5 (5.987176033927486, 0.634665266104897, 0.711055553385247)
0.5 0.1 (-1.9690929118423788, 0.7103520254071215, 0.31843366656181304)
0.5 0.2 (1.0023430870343295, 0.707530918052349, 0.44654227123532203)
0.5 0.3 (2.6995947405835983, 0.7142828571371427, 0.5480875842417887)
0.5 0.4 (4.139592830567267, 0.7037044834303671, 0.6373382147651275)
0.5 0.5 (4.885728516741401, 0.707530918052349, 0.6982836100038435)

[ZihaoCui]

or just annotatig the line 166, 167

[bngcode]

I came to the same conclusion. The way how th SNR is computed is resulting in
snr + 20log(|rmsnoise|/|rmsclean|), when using

noisescalar = rmsclean / (10**(snr/20)) / (rmsnoise+EPS).