Audio data collection scaling

I collected some experimental data using an industrial IEPE microphone. When I called audiowrite to write the data to a file, I received a warning message about clipping the data. I found a solution to scale the data by dividing by the maximum of the absolute value.

Looking at the data now, I did not realize the maximum would vary by so much since the measurements were taken at similar conditions. Since I do not have the original maximums I cannot properly scale the data to be on the same scale. Two audio files recorded at identical conditions should have roughly the same amplitude, but one of the files has a significant amplitude shift.

I was wondering if there are any techniques to scale the data properly. It seems that outliers are causing one of the signals to be shifted more than the other.

The regions from 0 seconds to 1000 seconds in the first signal and 200 seconds to 1000 seconds should have roughly the same amplitudes.

Do I need to recollect data?

Answer 1

[Walter Roberson]

Unfortunately when you normalize according to maximum absolute magnitude, then you can end up normalizing according to the maximum "pop" (or spurious noise), which can throw everything off.

If you have burst noise, then sometimes what you should do is a low-pass filter set to reject high frequencies (pops are fast, so high frequency); or sometimes use a median filter (which also tends to be a low-pass filter.) . And then, having done so, re-normalize each segment.

I see that those kinds of microphones are typically used to measure force or acceleration or pressure. I have to wonder whether using relative values between the measurements is meaningful, or if your work should really be using absolute measurements?

I could see relative being useful to study the distribution of vibration modes, where the interesting information might be in the locations rather than in the magnitude... but force and acceleration especially, absolute measurements are more often needed.

[Ivan Arkhipov]

Absolute measurements would be better but I could still make use of data on the same scale. I did have the max sound pressure recorded for the last experiment run, but I did not expect it to vary by so much. I am trying to compare different recordings for process monitoring. I used a median filter and re-normalized and it seems to do what I wanted.

[Walter Roberson]

How wide of a window did you end up using on the median filter?

[Ivan Arkhipov]

I used a window of length 10, but I am trying out different sizes. It seems that the median filtering and renormalizing works well depending on how much spurious noise there is in the segment I chose. For the files shown above, I applied the median filter to the entire file, and then chose 200 to 400 seconds as the region of interest to renormalize. I think I have to find the optimal region in each file for this method to be effective.