Problem Description:

I am splitting an input audio file(wav format) into 32 subbands using a QMF Analysis filter bank and again reconstructing it using QMF synthesis filter bank. I want to find how the number of filter co-efficients effect this process. To do this i have built matlab/simulink models with number of filter co-efficients being 20,40,20,80,100. I have passed a wav file through these and also have stored the output wav files.

I want know what is the metric that can be used to measure the distortion introduced due to this filetring. I am looking specifically for an objective value that can quantify the distortion



would be greatfull for all the help

If the output of your filtering is “time-accurate” you may use difference level metrics (http://www.soundexpert.info/articles/DiffLevel_AES118.pdf). Or you may estimate the difference between magnitudes of both signals – input and output. Any input signals including pure tones are allowed.

What do you define as distortion?

If you're doing a tree QMF the most accurate (LMS) you'll get is a 2-tap filter. (of course, only if you do no processing in the middle)

What kind of distortion are you trying to avoid?

If it's a PQMF are you worrying about the two-band-out aliasing?

What do you want to measure? What do you mean by "distortion"?

It could be

Frequency ripple
Far-band aliasing
Quantization noise in the filterbank

Heck, it could be just about anything.

If you put in a white allpass sequence, capture the output from it, and divide the transform of the output by the transform of the allpass sequence, and then take the inverse transform of that, you can get the impulse response that results.

What ever do you want?

You might want to try EAQUAL:

http://wiki.hydrogenaudio.org/index.php?title=EAQUAL

A couple of AES papers have used it as a metric for comparing quality when listening tests are not practical.

EAQUAL is very slow and not accurate enough. But that is the best that is publicly available.