Cleaning up a webcast recording, two puzzles

[Jonathan Sachs]

I'm trying to use DC6 to clean up a recording of an interview that I captured from a webcast. I'm looking for advice on a couple of problems that puzzle me.

I have prepared a zip file which contains two sound samples, each a few seconds long. It is too big to upload (about a megabyte), so I am posting it at http://home.earthlink.net/~jonsachs/extract.zip.

Sample 1 illustrates background noise which occurs throughout the entire recording. (It appears to have been present in the source, since it disappeared when the source was broadcasting recorded announcements, etc.) I thought the Continuous Noise Filter would be the proper tool for eliminating this noise, but when I followed the instructions in the help file, I had very little luck with it. I took a noise sample from the noise in the silent period in the middle of the sample file, then ran the filter, but it did not noticeably attenuate the noise until I pushed attenuation up to about 90. Even there the attenuation was marginal, and at higher settings the sound quality suffered. Playing with the attack and release times did not seem to have any effect.

I had better luck with the Dynamic Noise Filter, but I found its instructions harder to understand, and I don't really know what I'm doing with it yet. Before I invest time in learning to use it now, I want to ask how to judge which filter I should be using. If the Continuous Noise Filter, what am I being wrong that makes it fail to work?

Sample 2 illustrates a problem that comes and goes at some points in the file. Here it sounds like an echo; and other points it sounds more like "crosstalk" from another program. Can DC6 clean up this type of problem? If so, what should I try?

[Craig Maier]

I think that webcasts are transmitted via the use of lossy compression. That involves its compression via signal truncation in the frequency domain. That could cause an interaction between the bins used in the compression algorithm and the Diamond Cut CNF. What I would suggest trying is the varying of the FFT size. Try much smaller and/or much larger as I am not sure which way will get you away from the compression binning. Give that a try and let us know if that helps.

[Dan McDonald]

I got the first sample cleaned up pretty well with:

a) gain normalizing
b) using a multi-filter with
1) sampling from the first part (with no talking), with the settings at 125 attack, 250 release and 15 dB attenuation, and fft size of 8192 with 50% overlap
2) 2 notch filters (there's some bad hum) 1 at the 120 American hum setting and 1 at the 60 Hz American hum setting. If I wanted to clean the hum up even better than that, I would use the spectral filter.

Sounds pretty clean that way (except for the digital artifacts that are in a webcast).

Sample #2 is so nasty with echo - it cleans up fine, but then the echo is really bad. I've heard that fairly often on webcasts, but don't know what causes it. You could try playing with the l-r subtraction, but I've never had much luck with doing that. It sounds like about 3 echoes, so I don't know what else to do with that.

Dan

[Craig Maier]

The only system that I have seen to be somewhat effective with echo is the DSS (Dynamic Spectral Subtraction) Filter which is only found in the Live/Forensics version of the program.