My exported MP3 file levels are too high

I’ve been using Audacity for ten years and this is the first time I’ve encountered (noticed?) this problem. It came up because I’m re-mastering old podcast episodes for a new platform, which has a -2.2dB ceiling as a requirement for their audio files.
I have tried this on two Windows 10 machines, my laptop (Audacity 2.1.0) and my desktop (2.2.1).
The problem: I have a 30-minute recording of spoken audio in one track (mono) and another track of incidental music (stereo). I set my Normalization level to -3.0 dB. When I export the file to MP3 (Constant bit rate mode, 192kbps, Joint Stereo), the MP3 file is somehow normalized to 0.0 dB. I have verified this by re-importing the MP3 file to look at the waveform in Audacity, and also in Adobe Audition.
Is there a setting I’m missing?

Just for grins, does it still behave badly if you convert everything to stereo? This would be a condition odd enough that nobody would have checked it in development.

Does it still do it if you carefully selected each track and applied Normalization one at a time?

Koz

It’s normal for MP3 compression to boost some peaks, although a 3dB change is unusually high. As you probably know, MP3 is lossy compression, so the wave shape changes making some peaks higher and some lower. The new highest-peaks will often be higher, especially if the original file was dynamically compressed/limited.

It’s generally unpredictable, but since you’re getting +3dB you can normalize to about -5.2dB (or less) and after MP3 compression you should get -2.2dB (or less).

Some further info:
When I try exporting to a WAV, the same thing happens, it’s normalized to 0.0 dB.
I’ve experimented with the normalization settings. Going as low as -7.0 dB seems to produce an MP3 file where the variance between the original and the MP3 waveforms is visible.

Another thing I’m seeing is that after -7 dB normalization, is that the average level is now too low, but the peaks are now too high. I’m going to try using the Leveller and see what effect that has.

I also tried converting the mono track to stereo, then normalizing it. But when I export, I get the same problem.

It’s the first time that you have noticed.

As DVDdoug wrote, this is normal for MP3 because it is a lossy format, which means that it is trying to approximate the original audio but with less data. This approximation will frequently cause the peak level of the output to be a little higher than the original. The effect tends to be most noticeable with audio that has been limited, highly compressed, or clipped.

Below is a screenshot of a square wave. The top track is a WAV file, the second track is an MP3, and the third is an Ogg file. Note that of these three formats, only WAV retains the original shape of the square wave.

When I try exporting to a WAV, the same thing happens, it’s normalized to 0.0 dB.

There’s probably something wrong with your experiment. Regular (integer) WAV is hard-limited to 0dB, but Audacity uses floating point and has virtually no upper (or lower) limit. So exporting to WAV can bring the peaks down to 0dB, but they should never go up and they shouldn’t change at all if you are below 0dB.

I’ve experimented with the normalization settings. Going as low as -7.0 dB seems to produce an MP3 file where the variance between the original and the MP3 waveforms is visible.

That’s not surprising. -6dB is 50% and -7dB is less than that. (And then you said MP3 compression is adding-back +3dB.)

Another thing I’m seeing is that after -7 dB normalization, is that the average level is now too low, but the peaks are now too high.

The ratio (difference*) between peak and average is constant with linear volume adjustments… If you reduce the average by -6dB, the peaks will also be reduced by -6dB. Of course that’s before MP3 compression which tends to increase the difference.

You can try leveling, limiting, or the compressor effect.

IMO - The limiter set to Hard Limit is usually the most straightforward way to reduce the peaks with minimal effect on the average level and with minimal side-effects. It may take some experimentation/iteration to hit your targets with MP3.



\

  • Since dB are logarithmic, the difference (subtraction) is a ratio.

The solution I came up with: Rather than running a bazillion experiments to get each of 62 episodes tweaked to the point its MP3 files meets specs, I simply got a copy of Adobe Audition and edited the MP3 files directly. I spent way too many hours messing around with this, so mostly I’m still shaking my head.
Direct editing/analysis/processing of MP3 files is definitely functionality Audacity should include.