Initial DC offset in output files from Spleeter 1.5.1

[boy1dr]

Description
Using Version 1.5.1 of Spleeter, the output files generated using 2, 4 & 5 stem modes begin with a DC offset. The result of which is a pop when the output audio is played back.
Spleeter V1.4.3 does not exhibit this issue.

Step to reproduce
Install Python3.7
Run "python -m pip install spleeter"
Run "python -W ignore -m spleeter separate -i song.mp3 -o c:\temp -p 2stems.json"
No errors observed.
Check output files using audacity, files begin with DC Offset.

Output
Processing C:\Users\chris\Desktop\original.mp3
INFO:spleeter:File C:\temp\original/accompaniment.wav written succesfully
INFO:spleeter:File C:\temp\original/vocals.wav written succesfully

Environment
OS: Windows10 pro 64bit
Installation type: Python3.7 (pip) (tensorflowCPU)
RAM available: 32GB
Hardware spec: AMD Ryzen 5 3600 / GTX1660Ti

Additional context
The JSON config file can be viewed here
https://makenweb.com/2stems.json
An image of the DC Offset problem can be viewed here
https://makenweb.com/spleeter_1.5.1_dc_offset.png

[mmoussallam]

Hi @boy1dr ,

Indeed I can reproduce with the Librosa backend (used by default in 1.5.1). You can circumvent this by adding -B tensorflow in your command until we fix it. Thanks for pointing that out.

[mmoussallam]

Hi @boy1dr

We've made some changes that should have fixed this issue. Let us know if you still encounter it.

[boy1dr]

@mmoussallam Version 2.0 is clean as, seriously good work :)
Thanks for letting me know!