wind noise + authoring of disks

[VJ]

Hello,

My brother recently purchased a Sony DCR-TRV33 camcorder. Some initial recordings revealed that wind-noise is a problem (most of you will know this already ). Are there algorithms to filter it out as good as possible (I imagine the only high quality option is to use an external mic with windcover) ?

He is not that much into computers, and was wondering if there are stand alone devices that could provide usefull for editing (basic: cutting segments and perhaps adding titles) and recording to DVD. Along these lines, I figured there are several stand-alone recorders, and I think some offer (minor) editing possibilities (e.g. some Panasonics have capture from firewire + internal harddisk). Any suggestions ?


Jörg

[wkulecz]

For "free" try audacity audio editor, a Linux based project that has a windows port.

Last time I looked, it had a "spectral subtraction" (Weiner) filter where you choose a "quiet" section of the audio track and it does a spectral analysis and attempts to remove this from the remainder of the track.

The state of the are in this sort of thing is the Sonic Foundry noise reduction plugin, but its about $300.

--wally.

[VJ]

Thanks, I'll try them!
Not sure if the spectral subtraction will do the job, as the wind noise is changing constantly; but it can't hurt to try...

Jörg

[wkulecz]

You may have to break your audio into pieces at each place there is a "slience" and then process each piece seperately to deal with a changing noise. Tedious for sure, its always best to fix the audio or lighting before you record instead of trying to fix it later in software.

--wally.

[VJ]

Hmm, that is going to be tricky... There is quite a lot of wind noise (seaside), and the camera wasn't stationary. Still, might be interesting to try...

Is the windnoise traditionaly situated in certain frequency bands (perhaps lowering those might help; if it does not interfere with human conversation) ?


Jörg

[wkulecz]

That's basically what the Wiener filter does. It measures the spectrum of the noise and "subtracts" it from the spectrum of the signal. Frequency bands where there is only noise get quiet. Where there is speech the noise is "masked" by the speech. Its simultaneously done in many more (and narrower) frequency "bands" than you could do with an equilizer.

I found Audacity did a pretty good job of removing the noise of a servo motor from a sound track. This wasn't a constant noise either, but had a fairly constant "spectrum" over the duration of the clip. It wasn't perfect, but it was much less distracting than on the original recording.

Good luck!
--wally.

[guylauten]

You can try making a slightly delayed version of the audio and mix it in with the undelayed audio. This creates a comb filter response, and as you vary the delay, you vary the response. The rest of the audio may sound like it's at the bottom of a garbage can, but it can greatly increase intelligibility by "phasing-out" upper-mid to high frequency noise. Old tape deck trick.

[VJ]

Originally posted by wkulecz
That's basically what the Wiener filter does. It measures the spectrum of the noise and "subtracts" it from the spectrum of the signal. Frequency bands where there is only noise get quiet. Where there is speech the noise is "masked" by the speech. Its simultaneously done in many more (and narrower) frequency "bands" than you could do with an equilizer.

Ah, I see...

The speech isn't always sufficient to mask the windnoise, so it could well be that the "improved" version would sound very ackward... (only one way to find out... )

guylauten:
I'm a bit unclear as to why mixing with a delayed channel would work. Does one have to inverte the delayed channel ?


Jörg

[guylauten]

VJ asks:

"I'm a bit unclear as to why mixing with a delayed channel would work. Does one have to inverte the delayed channel ?"

If you mix an undelayed 1kHz wave with a 1mS delayed 1kHz wave, the delayed wave is presented to the mixer 180deg out of phase from the undelayed wave (the wavelength of a 1kHz wave being 1mS) , resulting in no output from the mixer. You also get this cancelling effect for all frequencies that are Integral multiples of the fundamental (in this case 1kHz). So the frequency response of such a filter would have very sharp notches at 1kHz, 2kHz, 3kHz, 4kHz, etc., making it look like a "comb" on a spectrum analyzer. As you vary the delay, you change where these notches occur. It can be very useful when trying to get rid of noise with an overall "white" or "pink" spectrum (like wind noise) while still keeping enough information for the speech to be intelligible. For the math, I refer you to here:

http://ccrma-www.stanford.edu/~jos/waveguide/Feedforward_Comb_Filters.html

[VJ]

Ah, it is becoming clearer...

I take it the delay has to be dependant on the audio-samplerate ?


Jörg

[guylauten]

Originally posted by VJ
Ah, it is becoming clearer...

I take it the delay has to be dependant on the audio-samplerate ?


Jörg

The delay is independent of the sample rate. It's easier if you look at it in the Analog/Frequency domain. It's about the relationship between the delay and the signal's wavelength. When you implement a comb filter in the Digital/Time domain things do work out better if your sample rate is well over 10 times the inverse of the delay time.

[wkulecz]

This may have been useful in the analog days when a variable delay could be created by changing the distance between a pair of heads in the tape transport stream.

Given that you have digital data to start with, I'd bet the Wiener filter would do a much better job in general. Not to mention that the analog equipment to do this is probably "priceless antiques" :-)

The "delay" needs to be an interger, N such that: SampleRate X N = delay

At the 48KHz sampling rate of DV audio, each sample is about 21 microSeconds apart meaning N would need to be 48 samples for a 1Khz "comb" (1 millisecond delay).

I've no idea why one might expect a 1KHz comb to do anything useful as opposed to a 1.5Khz comb or a 500 Hz comb. How many parameter sets would you want to try before you gave up or decided you've done the best you can?

OTOH with the old analog set up you could quickly twiddle the two gains and delay parameter while listening to the result to quickly arrive at the "best" improvement.

--wally.