[This thread was taken from personal e-mail correspondence I had with Paavo/Har-Bal on 11/8/04. The information contained in it (I think) is important enough to be shared with HarBal users. HarBal is a fantastic product, but users should be aware of possible issues when EQ'ing bass frequency content, as discussed below. If management does not agree that this information is worth sharing, it may of course delete this post!]
- john
zgraf@yahoo.com
================
Greetings,
I've purchased Har-Bal, and I'm very happy with it.
However, I've noticed (what seems to me like) some unusual processing in the bass frequencies, and I'd like to get your comments.
What I'm doing is taking a 24-bit WAV file, "smoothing out" the peaks and
valleys across the board (20-20K) with the parametric EQ, then pressing the "Record" button to save the processed WAV. Then I use Har-Bal to open the processed (filename_eq.WAV) file.
When I compare the newly processed WAV with the original WAV (as a
reference), I can see that Har-Bal did a splendid job of smoothing out the
peaks and valleys above (say) 80-100 Hz on all three waveforms. But
somehow, below the 80-100 Hz band things look a little dirtier. It would
appear that the Harbal algorithm isn't quite as good for smoothing out the
low bass region (e.g., < 80 Hz). In fact, sometimes my attempts to smooth waveforms in the low bass area actually seem to introduce more "abrupt" peaks/valleys and transitions in this region!
Can you comment on this?
Is it a reasonable thing to use Har-Bal to analyze a file that's already
been "Harbal processed", like I'm doing?
If so, shouldn't the average, peak, and mean curves in the "processed" (_eq) file look similar to the way they looked in the original WAV right before I pressed the "Record" button? [If the answer to this last question is "yes", then I'm observing some unexpected behavior in the low-bass region, as noted -- because there are discrepancies].
Aside from this one small issue, congratulations on a fine product!
John Jakob
zgraf@yahoo.com
==============
John,
The behaviour you see is normal and is to be expected. The frequency resolution of an FIR filter is dictated by its length in time. HarBal uses an 8192 point FIR which gives it a resolution in the
order of 5-10Hz at 44.1kHz sampling rate. As you increase the sampling rate this resolution will be proportionately scaled (ie. at 88.2kHz the resolution will be 10-20Hz). What this means is that if
you have two peaks that are 10Hz apart and you design a filter to smooth it out the realized filter will not be able to cos the peaks are too close together and the realized filter smears the desired
frequency response out giving one smeared peak instead of two. Because the frequency separation of peaks at the high frequency end is so much bigger the matching between desire frequency response and
what was realized is almost perfect. At the low frequency end the resolution limitation results in a departure that increases with decreasing frequency. The only way to obtain a closer fit at the LF
end is by using a longer FIR. The current length was chosen as optimum for 44.1 & 48KHz sampling rates. At some stage we will look into the practicalities of scaling the FIR length with sampling rate
to keep the effective LF resolution constant.
Regards,
Paavo.
Question about Processing of Bass Frequencies...
Great! Glad to hear that you folks are open-minded...
Have been using the program more and more--can't imagine doing without it now. It's really a wonderful tool. Can't say enough good about it (to those folks that will listen)!
I'd just recommend to Har-Bal users that they use a bit of extra care when tweaking response curves in the very low bass regions -- for reasons as noted.
- john
Have been using the program more and more--can't imagine doing without it now. It's really a wonderful tool. Can't say enough good about it (to those folks that will listen)!
I'd just recommend to Har-Bal users that they use a bit of extra care when tweaking response curves in the very low bass regions -- for reasons as noted.
- john