(Maybe it's an old question, so if it is, maybe forward me to another discussion topic...)

I read here about many tests between various endcoders, starting from a PCM file to the encoded one. Then some ratings are exposed. But i can't understand how you check the differences between original\encoded files. I mean: i should listen both? And if i have no HI-FI equipment (only MID-FI)?

I usually check the result of the encoding with a spectrum analyzer, and here i saw the impossible... referring to the mp3s that was not encoded by me. You know, LAME applies various lowpass filters: more the cut frequency is high, more the quality is high. A 128kbps has a Cf of 16KHZ, a 192 one has about 18KHz Cf.

The question is: is good to base on spectrum data to validate the quality of a LAME enc mp3? I saw for example that a 192kbps mp3 has about the same cut-off of a 240kbps VBR. In this case there are differences between them (except for filesize). And i also found the parameter -k (no filtering, but seems to work only for 3.90.3) and the output spectrum is amazing (near to full scale!) but there are here distorsions??

And in LAME 3.97, a 320kbps has an horrible frequency response (i still can see a net "cut" on 16KHz)!

Some explanations are very appreciated!

Thanks!

LAME (and all other kinds of lossy MP3, AAC, AC3, DTS, ATRAC... compressors) is so called perceptual coder. It exploits certain facts about the nature of human audio perception. So, you cannot rely simply on spectrograms to evaluate its quality.

It doesn't matter how "ugly" it is. What is important is if someone is able to distinguish between the sound from the original file and the compressed file through an ABX likstening test.

You don't listen with your eyes. It's as simple as that, hence spectograms are entirely worthless.

If you really want to check the quality of an encoded audio track you have to conduct a blind ABX test of the MP3 vs. the source material and tell by yourself whether you're able to discern both files from each other. The goal of lossy encoding usually is to find a bitrate which is transparent to the human hearing, hence a quality that's indistinguishable from the source can safely be called a decent one, no matter what a spectrogram displays on the screen.

I for my part usually ABX an encoded sample with different quality levels, starting from a low-bitrate one (like ~48 kbps with AAC and Vorbis) up to one that sounds transparent to me. Then I encode a few further samples using the bitrate which was transparent in conjunction with the previous sample and ABX these vs. the source as well. If at least one of these samples sounds flawed, the same procedure is repeated at the next higher quality step, until none of the samples can be distinguished from the source anymore. Having found this transparency threshold, I mass-encode using the next higher quality step as a safety margin for the possibly more problematic files which weren't tested before.

And about the listening tests on these boards: People give their ratings depending on the perceived quality of the material they listened to before in a so-called ABC/HR test. The better the file sounds, the better the rating will be, e.g. if the tester can't tell a difference between the reference and the sample it's a 5 ("imperceptible").

There is also the fact that not all music sounds good at a give fixed bit rate; hence the reason when I compress my music using Mp3, I tend to have a greater bias towards using variable bit rates at a higher rate, in the case of lame, I use -q 0 which gives me the best of both words, good compression whilst at the same time, not sacrifice the sound quality.

That's all interesting, but is long to read (and translate), so i'll check out next days.

But my idea was that if the spectrum of an encoded sample is more similar to the original, it means that the encoded one has not so many data lost.
I thought that mp3 encoding apply a lowpass filter, in oder to not take care of the too high and useless frequencies, but in fact it seems to use other criteria to perform the encoding, which i don't know.

I don't have a great audio equipment, so for me the faster way to compare mp3s was a fast spectrum analisys (a 128k looks trimmed and dirty, a 320k much more clean!)...

Have a look at this thread.



Have a look at the thread I linked above. Reading that should disabuse you of the notion. But here are a few choice quotes:

• ... you wouldn't dream of using your ears to decide on the relative merits of the pictures produced using different digital camers would you? ...
• ... Any chance somebody can get a Jpeg at 2 different quality settings, somehow convert them into an audio file of some description, then ask them which picture looks better based on the sound?
• ... The reason why Blade (nowadays an obsolete MP3 encoder) is described as "almost perfect" is the lack of a proper psychoacoustic model. Blade (and Shine, too) just encodes without exploiting the characteristics of human hearing. Vorbis on the other side probably does just that, what explains the somewhat ugly frequency graph.

There might be the nub of the matter. It's tempting to look for a fast fix. There isn't one.



But, as someone pointed out in that thread, Vorbis - which is a good codec that performs well in blind listening tests - doesn't look "clean" at all. Unfortunately, there's no short cut involving looking at pictures.

Nick E, I don't have much experience with Vorbis but when I looked the few abr 160 kbps ogg files I have on a spectrum analyzer, they looked "clean" (what I mean is, unquantized and close to how the original wav looked), much cleaner than mp3. However I think that is because the spectrum analyzer I used is geared towards looking at mp3s (its a 576-point MDCT, I think...) not Vorbis. Probably if I used a different analyzer with a different type of transform or a different freq resolution to match Vorbis (not sure what Vorbis uses, or what prog I should use to view) I might be able to see its quantization more clearly. Anyway, just wanted to point out that a particular encoding format might look totally different on different spectrum analyzers, depending on choice of freq domain, transform method, etc.

Jake, I agree with the majority thinking here that listening is much more important than looking at a spectrum analyzer, but I disagree with the majority thinking that the spectrum analyzer is completely worthless. I still think it has some worth, but in the end IF one can confirm that an encoded file sounds fine, then it should probably be deemed fine, regardless of how it looks. But in practice I think it is not so easy to confirm such things, and spectrum analyzers are pretty convenient so I like to look at them too.

> And i also found the parameter -k (no filtering, but seems to work only for 3.90.3) and the output spectrum is amazing (near to full scale!) but there are here distorsions??
> And in LAME 3.97, a 320kbps has an horrible frequency response (i still can see a net "cut" on 16KHz)!

In LAME 3.94 through 3.96, you must specify --noath in addition to -k to get full freq response. That's what I do. Most people here think that's idiotic, though. In LAME 3.97 you can fix that net "cut" at 16 kHz (it's not really a cut, but rather LAME 3.97 aggressively quantizes freqs above 16 kHz and I think that is what we both see) by NOT specifying -h (same as -q 2) on the commandline. You must specify -q 3 or higher (or don't specify anything is fine as well, I think) and that problem goes away. But there will still be a cut at 20 kHz remaining and there is no way to get rid of it in LAME 3.97 I think (one of the reasons I don't use 3.97).