Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: A question about MP3 and Vector Quantization  (Read 2967 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

A question about MP3 and Vector Quantization

This is posted on behalf of a new member the_g. This was asked of me via PM but as I lack the knowledge to provide an answer, I put it to the wider and more knowledgeable audience. ;)

Quote
Hello,
I'm new in this forum and I can't understand how it works. I would like to post this question but I don't know how to do it. I hope you can help me (i've found you randomly).
My problem is about the MP3 technique and the Vector Quantization lossy technique. I have a .wav file that I want to compress: for this I've written a matlab code and I've got some results. The fact is that: the values of the MSE between the original .wav file and the one compressed with VQ technique is lower and the MSE between the original .wav file and the one compressed with the .mp3 technique. How is it possible?
In this way it seems that the SNR of the .mp3 audio is lower than the one obtained with the VQ technique: but in reality the quality of the .mp3 file is way better.

I hope you can help me.. and sorry for the trouble.


Re: A question about MP3 and Vector Quantization

Reply #1
Hello everybody.
My question is this one: I have a .wav audio file that I compress with two techniques: an MP3 encoder and with the VQ technique. Is it possible that I obtain a lower MSE using the VQ technique? The MP3 technique should be more efficient I guess...



Re: A question about MP3 and Vector Quantization

Reply #2
MP3 uses a psychoacoustic model to decide which parts of the signal are important and which parts are not. The important parts of the signal are encoded more accurately (with less noise), and the unimportant parts are encoded less accurately (with more noise).

SNR does not have a psychoacoustic model, so it does not accurately measure which signal will sound better.

Re: A question about MP3 and Vector Quantization

Reply #3
Hello everybody.
My question is this one: I have a .wav audio file that I compress with two techniques: an MP3 encoder and with the VQ technique. Is it possible that I obtain a lower MSE using the VQ technique?

Yes, if you pick a quantization strategy that optimizes MSE rather then audio quality you will get a file that is badly compressed but that has low MSE.  Usually people don't choose to do this though since the goal of compression is to have good audio, not low MSE.

The MP3 technique should be more efficient I guess...

Depends on how you define "more efficient".  Audio quality is usually assessed by listening to the audio and evaluating how it sounds since simple error metrics don't capture how hearing works.

Re: A question about MP3 and Vector Quantization

Reply #4
Thank you very much.
So it means that I cannot do a comparison between VQ and MP3 since the first one has a lower MSE but the quality of the compressed audio is not as good as the one obtained with the MP3 (that has an higher MSE but a definitely better quality due to the psychoacoustic model).
I'm trying to make a Rate-Distortion plot between these two compression techniques but I guess it's useless...
Is there any way to plot a comparison between these two techniques?

Re: A question about MP3 and Vector Quantization

Reply #5
An ABC/HR test is commonly used to compare lossy codecs. There are many examples of these tests on the forums.

Re: A question about MP3 and Vector Quantization

Reply #6
A pure mathematical way to calculate it would require to implement a "Human-ear-emulator" (if you allow me the expression).
In other words, you would need to implement a psychoacoustic model that is an order of magnitude better than those used in the encoders tested.

And then, you would either need to plot lots analytic values ( better at stereo, better at cymbals, better at tonal, better at preecho better at....) or assume that the value that you will give will just be valid for the people that have the same judgement than the one you choose to reach that single value.
And that's without even looking at killer samples or different bitrate ranges.

That's why generally, listening tests with predefined conditions are done from time to time to actually get something similar to a ranking.

 

Re: A question about MP3 and Vector Quantization

Reply #7
Thank you all very much.
My last question is referred to the MP3 compression: in the VQ compression I had that by increasing the bitrate, the MSE decreased. Should the MSE decrease (if I increase the bitrate) also in the case that I use the MP3 compression?
I'm asking this because, I compressed online the .wav file in .mp3 with different bitrates and I noticed that by increasing the bitrate, the MSE was not decreasing. I think this could to the fact that:
1) The MP3 online compressor could not be so reliable;
2) In the MP3 compression, it's not said that by increasing the bitrate, the MSE decreases.

I hope you can help me. Thank you again.

Re: A question about MP3 and Vector Quantization

Reply #8
Should the MSE decrease (if I increase the bitrate) also in the case that I use the MP3 compression?

Probably depends on the sample, but on average I'd expect it to go down at least a little on most samples as you increase bitrate.

Re: A question about MP3 and Vector Quantization

Reply #9
That's what I was thinking too. I've done two tests with two different .wav files. I've compressed them in .Mp3 with Adobe Audition with different bitrates: 56kbps, 64kbps, 96kbps and 112kbps. With my Matlab code I computed the MSE between the original .wav file and the compressed ones: the results were that by increasing the bitrate the MSE increased as well in both files.
It's strange and I'm sure that my matlab code is correct. Any advise?

Re: A question about MP3 and Vector Quantization

Reply #10
Since you're looking at low bitrate files, it is probably due to how the encoder is changing various other settings as you adjust the bitrate.