[BYT][NOCODEC] alsa-bat check-alsabat-headset-playback.sh FAIL

[plbossart]

tracking https://sof-ci.01.org/linuxpr/PR2396/build4392/devicetest/

No idea why this happens. Our 'all green' status for CI devices lasted only a couple of hours.
@lgirdwood @lyakh FYI, not sure if anything changed at the firmware level?

[mengdonglin]

@singalsu Can you reproduce this failure locally?

[mengdonglin]

@plbossart @singalsu This seems to be the same random issue as thesofproject/sof#1988 that was closed without a fix.
Xiuli has checked the dumped WAV by CI and cannot see obvious wave distortion but cannot pass alsabat check. @aiChaoSONG will check the wav file with new CI tool.

[xiulipan]

@aiChaoSONG WAV files and pic uploaded. Please try some other measure method to confirm the error in the wav.
bat.wav.zip

[aiChaoSONG]

@keqiaozhang @mengdonglin @xiulipan @plbossart

Amp spectrum (whole wave, amp normalized to 1.0)

Amp spectrum(whole wave, local area)
side lobe peaks in dB: -27.356dB -23.496dB

Amp spectrum (exclude the first 20000 samples, ignore ramp, local area)
unit of frequency axis is frequency bin, so the peak located at about 1600 is not an error.

The ramp at the beginning of the recorded wave introduces higher side lobe, which may cause alsabat failed.
we can:

• add threshold in alsabat, make it less sensitive
• try to avoid the ramp at the very beginning
• implement wave analysis in wavetool.py, and use it for alsabat test.

[singalsu]

@aiChaoSONG It would be better to plot with dB y-axis, but I agree with your proposals. The ramp in the beginning impacts the spectra.

[plbossart]

@aiChaoSONG @singalsu can't we just skip the first part of the captured stream? if we are doing a frequency analysis it has to be done in steady-state, no?

[singalsu]

Yep exactly, the ramp part need to be excluded from analysis to get state. Also the FFT needs to be windowed if it's length is not synchronized to sine wave period. I'm usually using "kaiser(N, beta)" and using beta that gives suitable tradeoff between sidelobes attenuation and tone frequency bins spread. When applying window the level seen for sine wave is also impacted. Maybe start with beta = 10 for kaiser.

For playing with impact of windows and FFT length interactively I recommend to check this bat.wav file with sound editor called "snd" (from Ubuntu packages). Check from waveform the "f" checkbox on. Then go to transform options (single transform, dB, log freq, min dB 140, type Kaiser, size 1024, beta 0 - 1.0 (correspond to 0 - 15 as seen in plot), etc. If you select instead of single transform the sonogram, you'll see the distortion components of the ramp.

[aiChaoSONG]

I also checked wave file recorded by other platform, mainly JSL, the side lobe peak is not that high. So I am not sure if there is issue with BYT devices.

[singalsu]

@aiChaoSONG I edited my response above. To me the spectra in the bat.wav attached here looks quite OK. It's not 24 bit high-end quality but pretty good. There's some DC output above the noise floor (-100 dB or so, so very small, we may miss proper rounding somewhere in the test system and/or firmware but it's a minor issue).

[singalsu]

@aiChaoSONG The side lobes appear due to no windowing or insufficient windowing and due to different lenght of FFT than sine wave period (or integer multiple). E.g. the kaiser window with large enough beta is needed for >= 16 bit applications like SOF is. So it's a measurement technique issue.

[aiChaoSONG]

@singalsu In this case, alsabat is helping us to do wave analysis, after check its source code, I don't think it uses any window function.

https://github.com/alsa-project/alsa-utils/tree/master/bat

[singalsu]

Good finding @aiChaoSONG ! Computing of kaiser windows is somewhat tricky. Not sure any open-source library provide the needed functions. Does alsabat by chance use a fixed FFT length so the window could be pre-computed into a data file?

[aiChaoSONG]

Good finding @aiChaoSONG ! Computing of kaiser windows is somewhat tricky. Not sure any open-source library provide the needed functions. Does alsabat by chance use a fixed FFT length so the window could be pre-computed into a data file?

@singalsu Unfortunately, it doesn't, the FFT size it used is calculated by truncating time_length * sample_rate to 2^N, and time_length is given by user from command line.