Ever wondered, what happens to your music when encoding to mp3 ?
I always wondered, if mp3 not only removes psycho-acoustical "masked out" inaudible content (this is the common opinion about mp3 and all lossy encoding theories), but also
adds a fair amount of noise to the signal, which
was not yet measured or discussed.
So I measured the actual SNR (Signal to Noise Ratio) in the spectrum, and was very disappointed but confirmed about my listening observation.
I measured the SNR between busy frequency bands and silent frequency bands. The result is:
- mp3 128-256 constantly has only 25dB SNR, even mp3 320 has only around 32dB SNR
- mp2 (VCD, SVCD, Digital Video Broadcast) has only 10dB (160kbps), 18dB (224kbps)
- ac3 (DVD!) is not much better, 20dB SNR
- mp3pro totally screws the spectrum above 10kHz (removes bands, adds other bands or shifts bands in spectrum)
- musepack mpc is nearly as bad as mp2
- wav muLaw 8bit is not as bad as expected compared to the others (has 40dB)[]
- ogg (65dB) and aac (70dB) are winner in the lossy playground, wma9 follows (55dB)
- if you use an iPod, aac (or apple lossless (ALAC) of course) sounds better, many mobile phones play aac too.
- aac -HE (high efficiency) should be avoided, screws the spectrum just as mp3pro does. Use aac -LC (low complexity) instead.
- finally use the free lossless formats flac and ape to preserve your expensive sound and fidelity if you are scared from the pictures (as I am) []
How to read the pictures:
The first is the original wave spectrum. (equals to lossless ape and flac compression)
The peaks show the frequency response, the valleys show the noise floor.
Peaks should go straight up to 20 kHz, the valleys should be low and not narrow.
See yourself how different lossy compressed formats fill up the music with audible noise.
(http://mandala-in-mind.com/lossysnr/lin01.png)
(http://mandala-in-mind.com/lossysnr/lin12.png)
(http://mandala-in-mind.com/lossysnr/lin13.png)
(http://mandala-in-mind.com/lossysnr/lin14.png)
(http://mandala-in-mind.com/lossysnr/lin15.png)
(http://mandala-in-mind.com/lossysnr/lin16.png)
(http://mandala-in-mind.com/lossysnr/lin17.png)
(http://mandala-in-mind.com/lossysnr/lin18.png)
(http://mandala-in-mind.com/lossysnr/lin02.png)
(http://mandala-in-mind.com/lossysnr/lin03.png)
[/color]
(http://mandala-in-mind.com/lossysnr/lin04.png)
(http://mandala-in-mind.com/lossysnr/lin05.png)
(http://mandala-in-mind.com/lossysnr/lin06.png)
(http://mandala-in-mind.com/lossysnr/lin07.png)
(http://mandala-in-mind.com/lossysnr/lin08.png)
(http://mandala-in-mind.com/lossysnr/lin09.png)
(http://mandala-in-mind.com/lossysnr/lin10.png)
(http://mandala-in-mind.com/lossysnr/lin11.png)
(http://mandala-in-mind.com/lossysnr/lin19.png)
(http://mandala-in-mind.com/lossysnr/lin20.png)
(http://mandala-in-mind.com/lossysnr/lin21.png)
(http://mandala-in-mind.com/lossysnr/lin22.png)
(http://mandala-in-mind.com/lossysnr/lin23.png)
(http://mandala-in-mind.com/lossysnr/lin24.png)
(http://mandala-in-mind.com/lossysnr/lin25.png)
(http://mandala-in-mind.com/lossysnr/lin26.png)
This is very interesting to know for educational purposes - but the big question:
So I measured the actual SNR (Signal to Noise Ratio) in the spectrum, and was very disappointed but confirmed about my listening observation.
Can you also ABX the added noise with music(thus, non-testsignals)? If no, then it doesn't matter - because thats the purpose of lossy encoders: remove stuff which you cannot hear, and add noise/inaccuracies in areas where you won't notice them.
If you cannot, then you should remove the latter claim in your post, and add a notice that the added noise is inaudible to you.
....
Did you notice your valleys go down to 89 db? Maybe if you tried with a signal at -89 / 0db, you'd notice there's no problem?