posted 07 July 2001 04:28            

whitenoiseremoval.kym

I have been thinking all week about how to remove extreme amounts of noise from recordings and live shortwave radio reception. The difficulty was getting the spectral components of voice without unduly bending its spectrum. The autocorrelation filters failed in that regard because they multiply the incoming spectrum by itself. Hence, they overemphasize the bass and diminish the treble.

It finally occurred to me this evening, that we want to suppress signal components at the level of noise and leave the spectral line components alone. One way to do that is to use a frequency selective expander. In this case 512 bands. Signal amplitudes below the threshold get increasingly attenuated the further they are from that threshold -- the opposite of gain compression.

The sample has Kurt's Alien Threat buried in white noise (the toughest case to handle). When you try it, you will clearly hear Kurt speaking, but there is a lot of "granular" type gurgling. You can adjust the threshold and expansion ratio for maximum intelligibility. But then toggle the NR box and listen to what it sounds like without the noise reduction!

Those gurgles are the peaks of the noise occaisionally poking up high enough to get amplified. The granular type sound occurs for the same reason that granular synthesis sounds the way it does... little grains of sound lasting for very short times. The duration of the grains here is 1/2 FFT interval, or about 5 ms with 1024 FFT cells.

Now under more benign circumstances, like removing noise from tape and LP's the noise is already many dB below the signal, and this technique works wonderfully well. Simply adjust the threshold down into the signal level so that it get unaffected. Anything above threshold is given unity gain. The noise lying beneath the threshold gets its gain expanded downward so that it becomes weaker.

If your noise background isn't white then simply make the threshold into a frequency dependent lookup table (waveshaper).

I don't know for sure, but I'll bet this is how those noise removal programs work.

Cheers!

posted 07 July 2001 05:44            

man, youīre lightyear ahead of me in kyma programming and dsp-knowledge in general. iīm allways impressed about what you "squeeze" out of the capy

i had to use the new DNS1000 denoiser made by cedar the other day. the fascinating difference to their stand-alone de-hissers and de-noisers is that you donīt get any digital sounding artifacts, no matter how radical your settings are. even removal of camera noise from dialogue is quite easy (a problem that was knows as unsolvable in the past). the machine has no latecy (around a few mikroseconds).
in one word: itīs great and you donīt get the "robot" voices like DINR or all the other denoising programms make.
i wonder hoe they do it.
the cedar sound more like the old dolby Cat. 430 module wich basically is based on a dolby SR expander module but gives you access to the thresholds of the bands.

posted 07 July 2001 11:47            

whitenoiseremoval.kym

Attached is the original plus an improvement for the situations with extreme noise like I am facing. While the original had the gurglies, this new one gets rid of most of those. It uses a little bit of incoherent autocorrelation on the incoming spectral magnitudes before going into the expander bank.

Try it with Kurt's threat now -- sitting at 1:1 with the strong white noise background. There is a parameter !NInt that indicates the number of FFT subframes to smooth over using an exponential averager. The longer you make this the more it sounds like there is a slight echo. I keep it around 10 frames and the echo is not too bad. Remember that here I am concerned merely with understanding speech that is buried in the noise -- not hi fi reproduction.

I don't know how this will behave on your noise removal projects. It might be a bit more robot voice like. I'm not sure what that term means, but I have seen others talk about that effect too.

[Bear in mind, that if you have recordings then you get a chance to peek at the future and you should be able to do even better!

Generally, if you have a sample of noise-only, then you should be able to improve this significantly by shaping the threshold to follow the noise spectral shape, and setting its values equal to the noise level at each frequency.

I'm working blind about the future when I listen to the radio live. I don't even know what the noise spectral shape is... In that case this white noise assumption seems to handle most cases pretty well, and I adjust the threshold to above the signal level so all of it gets treated to some extent.]

Cheers!

- DM

posted 07 July 2001 13:25            

It works better than the noise removal plugins that I had in my sound editing environment! Wow!!

I don't know if this is how they do it, but this works incredibly well! Even without special threshold shaping!

- DM

posted 07 July 2001 13:58            

I just checked on the Cedar DNS1000 denoiser... It looks like they are using dual SHARC DSP's. I am curious about how they can pull this off with only a handful of samples latency. That isn't even enough time to estimate frequencies.

Does the Cedar take a little bit of time at startup to train itself? (Looking for hints that it might be an adaptive Widrow type LMS FIR system). Do you have to feed it some samples of noise alone?

Were you saying above that the *Cedar* does a terrific job of noise removal? Or were you implying that my Kyma algorithm does a good job? Not sure how to read your statement...

- DM

posted 09 July 2001 10:49            

I tried running it on a Capybara-66, but ran out of static RAM...can anything be done?

posted 09 July 2001 10:51            

It sounds like the FFT size is too large for your available memory. Try dropping the FFT size parameter in the rightmost sound which is a script.

Cheers,

posted 12 July 2001 18:11            

just coming home from a 6 week trip to the states,
mathis

posted 13 July 2001 02:00            

i think on dialogue it pretty much sounds like someone traying to speak while his mouth is full of water

posted 02 July 2002 03:34            

could you please post your sound again? i would love to try it out on my current project and lost it.

posted 02 July 2002 12:10            

quote:

---

Originally posted by David McClain:
I don't know for sure, but I'll bet this is how those noise removal programs work.

- DM

---

Indeed. I believe this is how both DINR and Ionizer work. Ionizer also
does variable band-to-band corelation to reduce gurgling.

posted 03 July 2002 09:27            

it seems that david also deleted these files. (that was long before you extended file upload possibilities.) could you please also upload these wavetables?

posted 15 July 2002 23:58            

quote:

---

Originally posted by Bill Meadows:
Indeed. I believe this is how both DINR and Ionizer work. Ionizer also
does variable band-to-band corelation to reduce gurgling.

---

Wow! Now I would be very interested to find out exactly what you mean by interband correlation. Can you shed any further light on this?

Cheers,

posted 16 July 2002 17:02            

quote:

---

Originally posted by David McClain:

Wow! Now I would be very interested to find out exactly what you mean by interband correlation. Can you shed any further light on this?

Cheers,

- DM

---