Hacker News Reader

Comment by cJ0th | Hacker News Reader

> Another thing for engineers or anybody at a station to do is to go into the studio, turn the microphone on, crank it to 11. Don’t talk — so you don’t blow your ears out — and listen to the sounds of all the fans that you have in the room. This is one challenge that we’ve had for decades here. You’re never going to get it to zero. You’re never going to get it completely silent. Nor do you really want to, because in order to get it silent you’ve got to move a lot of equipment out, and there’s a lot of cost with that.

Makes me wonder why they don't record the sound of the fans and then invert that audio track and mix it with the mic channel before broadcasting the master channel so that the relatively periodic noise of the fan cancels itself out.

View on HN

Because it's not really periodic enough. I don't think any natural sound source short of a musical instrument (and even for them, it's doubtful) is so perfectly periodic that an out of phase version would stay in opposition (i.e. in the inverted/opposite phase of the cycle) for long enough for this to be of any use. But there are (FFT-based?) noise removal tools that can work for this sort of thing.

edit: also, the weird phasing artefacts this would introduce would be much more prominent and noticeable, even if it succeeded in reducing the overall volume. So much so that doing basically this ('flanging') is a standard musical effect for making parts sound psychedelic.

Follow-up question, what if you put a separate microphone further away from the speaker to get a synchronized sample of the background noise to subtract away?

I suspect the logical end to that sort of thing is to use multiple microphones to add "depth" to your sound recording, then filter out sounds coming from sufficiently far away.

Like the way the Grateful Dead used a pair of microphones to cancel out feedback from the P.A.

It worked for them and would work, so long as the person you are recording keeps their head very still. Any movement would change the parts of the sound emerging from their throat that gets cancelled by the cancelling mic.

While it is indeed a cool idea, it really does create more problems than it solves-- there are reasons why it never caught on in live sound reinforcement.

That's how noise cancelling headphones work but it introduces audio artifacts that are often just as bad as the thing you're trying to remove.

The thing most people don't consider is that white noise (hiss, a fan) is readily tuned out by the brain - the brain is brilliant at reducing background noise - but artifacts introduced by artificial noise reduction are correlated to the signal and the brain interprets it as a distortion of the signal, not background noise.

That's why audiophile audio engineers fix the noise instead of using FFT noise reduction - at least whenever possible.

This is true that the brain will filter out steady state noise, but this effect is much less strong once a signal has been recorded.

Basically, on playback, the steady state noise sounds much louder than it did when you are actually present in the room.

The same holds for the reverberation in a room, it will sound much louder on a recording than on location.

For example, if you record someone speaking in a reverberant space like a church, it will sound like they are in a very reverberant space. Which they are, when you are listening live, your brain filters out the "non-direct" sounds so you notice just the voice.

I'm pretty sure these noise filtering processes are related to the high priority the brain puts on understanding speech.

All true.

However FM radio always has white noise, so adding a little bit to it is probably better than adding any correlated noise from signal processing.

FFT noise reduction definitely has its place and I do use it from time to time. I just highly doubt "live on-air signal processing" is a good application given that the noiseprint is constantly changing.

These broad generalizations about psychoacoustics are simply not accurate. It sounds like what every novice engineer experiences before honing his or her craft. If you're surprised by what you hear after recording something, either you weren't listening properly while recording it, you haven't learned how to record it properly yet, or your recording/listening environment itself is flawed.

Of course we notice different things about an object after taking a picture of it. But you've got to be very careful about cause/effect/bias.

I've been a professional recording engineer for 20 years. The phenomena I'm discussing is an objectively true, subjective perceptual effect, it doesn't change based on how long you've been recording. Over time you learn to expect it, but it doesn't go away. A lot of audio engineering involves working around the heavy real time signal processing our brain is always doing.

That's interesting about animals. It seems likely that their auditory systems would be tuned for the type of stimuli that are vital for their survival. There would be a lot of overlap, but since it's hard to tell their subjective experience, it would take some clever testing to see how their perception of recorded sound would differ.

Non-human animals have all the same "noise canceling hardware" as humans (It's mostly in the hind-brain), so it can't just be for speech. I know that that echo/noise canceling hardware plays a big role in sound localization.

This is probably why the noise and echoes sound louder on a recording. They don't change the way that they should when I move my head, so my brain can't figure out how to cancel them.

Having made a few records myself... when mixing with noisy source tracks, I generally choose to use console automation to mute the spaces between the notes, rather than trying to remove noise from actual signal, or even gate out the noise in the quiet spots. But it's more important to simply prevent noise in the first place.

On the first record I ever produced, I was credited with playing "furnace". We were recording in subzero January weather in Minnesota, and the furnace in my basement was very noisy. I'd go and turn it off, and we'd track as quickly as possible before the house got unacceptably cold and the furnace needed turned on again.

As an interesting aside, something a lot of newbie/hopeful audio engineers obsess about is "mic bleed", where the mic for one instrument picks up sounds from another. I mostly ignore this. Bleed is swamped by signal, and disappears into the mix. Even if a punch or edit fixes an error in the "bleed" instrument, it usually drowns out any error that remains in the bleed itself. Then again, audio engineering is full of a lot of nerding about second and third and even fourth-order effects, while ignoring first-order effects like "Is the song actually good?", "Is the performance good?"

> I generally choose to use console automation to mute the spaces between the notes.

I recently had to make a promo video for my startup on short notice, with a less-than-idea setup, and I did the exact opposite. I filled all of the 'silent' parts of the video with 'ambiance' because the sudden drop to absolute silence sounded very jarring.

I have no experience will audio engineering. I just thought it was interesting that we took opposite approaches to similar problems.

https://en.wikipedia.org/wiki/Ambience_(sound_recording)

Horses for courses. This is why a little background music can be helpful - it fills those silences. In the case of mixing a record, there are other instruments in the silences, so noise is just competing with music.

> something a lot of newbie/hopeful audio engineers obsess about is "mic bleed", where the mic for one instrument picks up sounds from another. I mostly ignore this.

A better approach to "removing" bleed is instead to make sure the bleed sounds good.

Euphonic bleed is solid gold.

So in all audio, there is "error". Error which is correlated with the signal is called distortion. Error which is not correlated with the signal is called noise.

Amplifiers use negative feedback to cancel distortion out ( because it's correlated ). Some folks, like at Pass Labs, tend to avoid neg. feedback for philosophical reasons, and people like their amps. They use part selection and careful design instead.

Adding two uncorrelated signals together, you end up with a louder, uncorrelated signal.

There is "single-ended noise reduction" ( the version I use is from CoolEdit 96 and CoolEdit 2000 ) but it leaves artifacts. It's still wonderful for things like single-coil hum, things with a limited spectrum.

Fan noise isn't even remotely periodic, because most of it is due to moving air rather than moving blades.

However, you can take advantage of that because the sounds you want to keep generally are periodic at short timescales. You can use DSP filters to remove components of the sound that don't show significant autocorrelation. The same idea lets you remove white noise and random static from radio signals.

That would be like trying to silence a fan by buying the exact same model and putting it on the opposite side of the room. It will just sound like twice as many fans.

Getting two sounds to perfectly interfere is basically impossible. Real world sounds are way too complex for that.

This particular technique has its problems, as has been mentioned by other people. I would think inverting and cancelling is much more useful in situations where you can isolate the background noise in real time. For example, I do a little hobby music production, and I've heard of this being used to isolate vocals:

Because the vocals are sometimes the only thing that's perfectly in the middle of the mix (balanced left/right), you can invert the left channel of a mix and then recombine it with the right to cancel out only the balanced tracks (so you're left with essentially an instrumental). Subtracting the instrumental track you've just created from the full song (again, by inverting and recombining), you should get the vocal track by itself. Fun stuff, right?

I'm not sure how well this works, as I've never done it myself, but it would seem like having the exact signal is important. As people have said, the fan noise isn't predictable enough to cancel using that technique. On the other hand...

I recently read about a machine learning application using K-means clustering in which you feed the algorithm sound from two microphones set up in the same room and it's able to separate the audio by who is speaking. I'm not sure how well such an algorithm would work when one sound is significantly louder than the other (such as quiet fans and people talking), but it certainly points to that being a possibility.

As a part time MC, I can attest that inverting and recombining channels of a song can occasionally lead to a decent instrumental. That being said, with little talent or background, and using freeware, it will not be perfect. It will be much better than using the original track, but it will certainly be noticeable that you are not using the instrumental. Great for practice though.

NMF and ICA can also be used for blind source separation, but there are artifacts. Good for some applications like speech recognition, but for music you really want no audible artifacts

Tools like iZotope RX can do a pretty great job at removing that kind of noise.

True, but a lot of the tools within RX are meant to be very manual labor-intensive, very CPU intensive, and not particularly realtime.

At the end of the day, the cleaner the source material, the easier it is to process in any fashion you want later on in the airchain. A lot of stations will have noise gates on their microphones, or even wideband gating before material hits the transmitter in some cases. With the right settings, it should be more than plenty.

I think audio fingerprinting for noise reduction can be automated but it will not real-time of course

All the noise reduction algorithms I've heard so far introduce fairly audible phasing artifacts. I think better noise reduction for a known/trainable source like fan noise could be done with machine learning - it's something a few folks are looking at

Most modern recording studios store all equipment in sound-proof closets.

The challenge is keeping those closets cool with ventilation that does not leak sound.

Well away from leaking sounds in to the room that the recording is done in. Even desks can have fans in though and they need to be in the same room as the operator.

Ask any DAW studio owner for a tour. The first thing they will show you is the isolation of all fans/equipment from recording booths and mixing areas.

It's a point of pride in studio design today. The NPR article mentions similar isolation from the newsroom.

Fans have a tonal center around 110-220Hz, then a lot of harmonics that essentially generate pink noise... nothing that can be corrected in post production or flipping the phase.

Or why not use a shotgun microphone? That's what we (well, audio guys) generally tend to use on film and TV sets.

Does not eliminate room noise. If you have worked sound on a film or TV set I'm sure you are familiar with capturing room noise on a quiet set to mix in during post.

Yeah, you're right. Sound guys always pick up room tone before or after the shoot. I was under the impression shotgun was used to eliminate/avoid sound sources that are not desirable. From my, limited sound experience, I've seen sound guys use it like that on my sets. For example, there's some noise from certain direction, but I am told not to worry about it since it won't be picked up with shotgun. Lo and behold, it wasn't there in the source material. It's probably due to many factors I'm not aware of though (sound reflecting of stuff etc.). They do turn off all the A/C's, fridges, and everything that's humming and buzzing though.

The reason shotgun microphones are used is because the polar pattern[1] is highly directionally. Basically the amount of sound that a shotgun type microphone picks up is very strong in one direction and less in another. This reduces peripheral sound being recorded at the same level as your target sound.

[1] http://www.proaudioland.com/wp/wp-content/uploads/2014/08/po...

Studios normally use microphones with larger diaphragms, because they sound fuller. But they're too fragile for field recording. Hence small-diaphragm condenser microphones, like your shotgun.

Shotgun mics have a pinched sound quality - the whole point of using U87s is the way they sound.

They can switch to SM7s or RE20s and achieve much better S/N with respect to air handling systems, but they want the "U87 sound" (they can keep it, too, IMO - it's painfully neutral) and that just entails a much higher "room noise to signal" ratio.

Probably be better to get some of the highest frequency components and then filter those out.

Reply on news.ycombinator.com