What's the big deal about latency?

[-]

Forgive me if this is a dumb question, but I was looking at the Sonar thread and reading about people's high latency problems and wondered...

 

What's the big deal about latency?

 

Yeah, the meters are a little late compared to the input signal going in, but what are some situations where there is a concrete advantage for having low latency?

 

My soundcard has a fairly high latency, yet for now I'm perfectly happy recording, overdubbing, and mixing with it. What am I missing?

[miroslav]

I had a latency thread going on David's forum with soame additional links included to other related theads...check it out.

 

http://www.musicplayer.com/ubb/Forum17/HTML/001086.html

[-]

I'm assuming latency is a problem for overdubbing... But if you either split the signal going into the card or, if your card drivers (like mine) automatically send back the immediate signal through the outputs anyway, then I still don't see how this is an issue.

 

When the input signal goes through my card, there is very, very, very little delay. You can hear slight phasing if you combine the thru signal with the original signal source, so there is a slight delay happening, but no more than a delay you get from plugging your guitar into the POD. (I send the signal directly into my card, so as not to get any crosstalk from other mixer channels.)

 

I skimmed through the links above... I understand why latency happens, but I don't understand why this is considered undesirable from a recording perspective.

 

 

 

This message has been edited by popmusic on 06-12-2001 at 04:01 PM

[Anderton]

In a nutshell, in a lot of cases it doesn't matter. It matters in two main areas: Playing softsynths from an external MIDI controllers, and monitoring.

 

For the first example, latency means that you hit a note on the keyboard, but don't hear it for several milliseconds later. This can be devastating if you're trying to play a drum part. I've found that as long as it's below 20 ms, I can cope; 10 ms feels a lot better, and under 5 ms is ideal.

 

If you want to monitor through your program (e.g., to hear how a plug-in affects the sound, or if you're using your computer in live performance as a signal processor), latency is again a major problem. For this reason, most musicians monitor the signal going in to the system so that there's no annoying delay.

[-]

Thanks, Craig -- that makes sense... I don't use softsynths and I don't use my DAW as an external signal processor, so I guess that's why I never found this to be a big deal.

 

Is there any DAW out there where with near zero latency? (Not that I need one... I'm just curious.) Or is that an impossibility given what we ask our DAWs to do?

[miroslav]

Well...I too am trying to get a handle on when latency IS a problem, though I've read in posts where some people are experiencing something equivalent to Slap Echo...which WOULD make it difficult to playback tracks and record new ones all in sync.

 

I guess every cpu/soundcard rig and your working habits are a bit different...going in doesn't seem to be a problem, but when you have say...8 recorded tracks in your computer, and now you want to play those back as a guide while you record your 9th, 10th, 11th...that is where any type of delay would be a REAL pain in the ass...

 

still checkin'

[-]

Originally posted by miroslav:

Well...I too am trying to get a handle on when latency IS a problem, though I've read in posts where some people are experiencing something equivalent to Slap Echo...which WOULD make it difficult to playback tracks and record new ones all in sync.

 

I think it would be impossible to work like that. I can get that effect with Cubase and my card (Audiomedia III) if I turn on the monitoring of the channel (within Cubase itself) as I'm recording.

 

If I need effects while recording (say I'm recording vocals and I want to hear some reverb while I'm singing), I don't rely on any plugin reverb because then I would get the delay.

 

What I'll do (The Audiomedia III is a 2-channel card, so this is a big time workaround) is pan the rough mix to the left channel and have the dry (live-in-the-process-of-recording) vocal being sent through the right channel. I'll pan the result of both channels in my analog mixer to the center. Then I'll use the aux send/returns on my analog mixer to get some stereo reverb on the right channel (the dry vocal channel). Obviously, the reverb doesn't get recorded -- it's just there in the mixer so I can hear it through headphones while singing.

 

Of course, it's an inconvenience to be essentially monitoring in mono for vocal overdubs, but it works fine for my needs.

 

 

 

This message has been edited by popmusic on 06-12-2001 at 04:44 PM

[Master Zap]

Latency was not a big issue at all for me before the day I started using virtual synths...and try to play them in realtime. As Craig says, even 30 milliseconds is... annoying. And the 300 milliseconds which was the default of my driver was... not fun http://www.musicplayer.com/ubb/smile.gif

 

 

Funny though how so many people misunderstand why latency exists.

 

 

In a hardware, say, a reverb unit working "inline", sound enters, get to a D/A converted and becomes a digital signal, but the same signal appears with only a sample or two's delay on the output. Any effects that are added are added "on the fly" by the unit. Since the unit is built for massive parallell processing of data, it all happens "at once", taking once sample at the time, and doing its magic on 'em.

 

 

In a computer it works differently. Latency stems from one thing - buffers.

 

When a sound sample arrives at your soundcard, there is a A/D converter there making it a digital number. Now it would be a royal pain in the rear if the A/D converted wanted the processors attention 44100 times per second, elbowing the poor processor and saying "hey processor, I gots me a new sound sample, you wanna?!".

 

So to avoid this, there is a buffer.... the sound card writes into memory, a bunch of samples (can be anything from a few hundred to thousands) and THEN elbows the processor and says "messa gots meslef a whooole fiine bunch of them samples, youssa wanna?" and the processor take's em.

 

This means that X number of samples has to come in before the computer even knows they are there and can do ANYTHIGN with them.

 

Now, the processor wants to do something with these samples, i.e. throw it into a Direct X plugin. Now you can do this with a small chunk... or a large chunk. Now, packing up the data in an appropriate chunk, and passing control to the plugin takes a little overhead in doing. The plugin, when it gets called, has a little setup work to do, before it processes the chunk.

 

Because of this, processing four thosand 10-byte chunks is QUITE a bit more work than processing ten 4000-byte chunks! But remember, that the computer cant begin to process the chunk before it even HAS it. So for a 4000 byte chunk, we would have to have read in all 4000 samples before we can even begin processing them. Assuming the processing itself is instantaneous (hardly the case), the best-case delay is still 4000 samples.

 

Hence, for low latency, one tries the four-thousand 10-byte chunk approach instead of the ten 4000-byte chunk approach.... or at least smaller, rather than larger, chunks.

 

So now this chunk has been thrown to a plugin who has done "its thing" with it. This processing has taken some finite amount of time. Maybe the chunk needs to be sent to another plugin.

 

Now, the chunk will have to be OUTPUT.

 

Now, soundcard OUTPUTS works sort of like the input. There is a buffer, which the soundcard automatically scans through and play at a regular pace of e.g. 44100 samples per second. When it's done with one buffer it nudges the procesor rather hard and says "oi, processor, I'm done with this here chunk, now get me another one PRONTO".

 

The processor better make damned sure that the NEXT chunk is ALREADY THERE, processed and DONE, so it can INSTANTANEOUSLY tell the soundcard (before even the timeframe of one sample has passed) "okay okay soundcard, you will continue reading from THIS position - tell me when you've read and output X number of bytes, okay?".

 

Soundcard grumbles, but agrees to do as he is told, and takes the new chunk and starts outputting is. Now the chunk which the soundcard was just DONE with is now available to us as an output buffer, so the processor will generally now copy some data to this buffer, so it can, the next time the processor puts its elbow firmly in the rib of the processor, have it nice and ready with new data.

 

Now, this copying might, if we are unlucky, take a little time... maybe it needed some data off the harddisk, and them thar harddisk controller was asleep AGAIN. ("Damn", the processor grumbles, "I wonder what that Segate piece of crap really IS doing, I asked for this block of data TWO DAMNED MILLISECONDS AGO!!").

 

So what if it's not done with this buffer, when it gets elbowed by the soundcard wanting a new block of data? Well, this is such a common occurance that in genral there is at least four rotating output buffers used, each of X samples in size. In theory two buffers would suffice, but hey, who of you guys live in Theory? I know, one should move to Theory coz there, everything works, but hey, I'm in Sweden, you?

 

So. That sound sample that came in the A/D has now had a journey over various buffers. Before it's "companion sample" (i.e. the processed version of itself) comes out the D/A the other end, TIME has passed. This time is known as latency.

 

HOWEVER. Processor is smart. Coz, processor figures, "if this junk AINT coming in through them thar INPUTS, but from that LAZY BUGGER SEAGATE" (processor eyes the harddisk with a huge frown) "I can COMPENSATE for this... since I *know* the latency is n buffers times X bytes, I can tell SEGATE this much time in ADVANCE what I want. And I'll be ready with it JUST IN TIME". The processor smiles - a brilliant case study of "Just in time" delievery, righ there in the heart of your DAW.

 

Thats why latency DOESNT matter for general multi-track work. The software compensates. Only when doing a straight-through-live processing situation (or playing softsynths, which is also a sorta straight-through-live processing, albeit from MIDI data into WAV data).

 

 

Hmm. Do I ramble? I do. But admit, it was a kinda neat description of latency, no? http://www.musicplayer.com/ubb/smile.gifhttp://www.musicplayer.com/ubb/eek.gifhttp://www.musicplayer.com/ubb/biggrin.gif

 

/Z

[Tim Walters]

An excellent and witty summary, except:

 

Originally posted by Master Zap:

In a hardware, say, a reverb unit working "inline", sound enters, get to a D/A converted and becomes a digital signal, but the same signal appears with only a sample or two's delay on the output.

 

Not quite. The D/A conversion itself takes a couple of milliseconds. Not much, but enough to cause problems if you mix the delayed signal with the original.

[Alndln]

Near zero?? The Rme Hammerfall(audio card) boasts 1.5 ms (asio)which is great,but not usable if you have a ton of audio tracks as well. In other words the more audio tracks you have and the lower your buffers are set the more your'e going to experience dropouts.Also the higher the sample rate(48,96 vs/ 44)the easier it will be to get away with the lowest possible setting.The trick is to get a usable balance.Personally I try to get all my soft synth stuff done first,or at least as early as possible so I can get low settings,then switch to between 85-100 ms for audio and the rest of the project.I'm sure a solution is looming on the horizon,or at least I hope so.

[Master Zap]

With all due respect,

 

you must have a severly BROKEN A/D converter if it takes milliseconds (thousanth of a second) for it to decode. And how on earth it is supposed to be able to do this 44,100 times per second is beyond me http://www.musicplayer.com/ubb/wink.gif

 

A/D converters normally take nanoseconds (billionth of a second) to do their task.

 

However, most units, even hardware rewiobs, really do a little buffering for various reasons, so yes, you can get a nasty phlango-effect when mixing the original with the effected if the effected has dry elements in it.

 

However, its still theoretically possible to get this time down to less than the lenght of a sample, if the unit is fully parallellized.

 

/Z

[miroslav]

Master Zap...enjoyed the blow-by-blow...you should think about getting a cartoon illustrator to add some comics to go with your funny explanation!!!

 

Anyway...it seems to me that at this point in time, a particular rig/working habit might or might not give you latency problems...and that there is always some latency.

 

No such thing as true "real-time" multitrack recording/editing/mixing in the CPU world...we just learn to live with the audio slur/blur...hey, it might actually become totally acceptable!

 

Look what "out-of-focus" imaging has done for the visual arts...in most cases, it is actually "more pleasing" than hard, crisp images...how's that go..."if it is out of focus it's art, and if it is in focus it's pornography"...hahaha!

 

So...anything under 5ms is becoming the "acceptable latency standard"...?

[hiraga]

Most converters actually do have a latency that's quite noticable. Hardware efx units too. Just try running one, converter or efx-box, in parallel with the original signal and it's quite noticable, and often makes it impossible to do various tricks like multing...

 

The Paris system from Ensoniq.. don't they brag 'bout "zero latency"? (I haven't tried a system, but from my understanding is a VERY good - as in VERY good sounding - system.. hmmm.. 'nother gizmo on the list, I guess..)

 

Funny, with todays digiblah/DAWs no-one monitors the RECORDED signal anymore??! Oh yeah I forgot.. digital is *perfect*.. hæhæhæh..

[Al Pickard]

Originally posted by Master Zap:

With all due respect,

 

you must have a severly BROKEN A/D converter if it takes milliseconds (thousanth of a second) for it to decode. And how on earth it is supposed to be able to do this 44,100 times per second is beyond me http://www.musicplayer.com/ubb/wink.gif

 

A/D converters normally take nanoseconds (billionth of a second) to do their task.

 

 

Most of todays ADC's are delta-sigma converters, which house a digital filter. This filter will have a delay of some fixed number of samples. For example, the Crystal CS5396 has a delay of 34 samples, or 0.7 mS at 48kHz sampling rate. This is completely dependent on the design of the digital filter.

 

Al Pickard

Digital Audio Labs, Inc.

[spigotsaol.com]

Just to chime in,

 

Latency is a huge factor when doing OD's with good players. As has been stated here before, some guys (or gals) can notice .500 ms in adjusting their feel. Try stacking 12 tracks of BKG Vocals with 12 singers on 3 mics with 2-5 ms of latency in the system, each time you record. Pass one is ever so slightly behind the track. The next stack is later still, because they are trying to match pass one, which was late to begin with, and so on. I get this kind of delay using Pro Tools Mix, an O2R (monitoring only) and a PSX-100. Anybody got any hints to lessen the delay? I've turned off ALL DSP that I can in the O2R that might add to the latency, and obviously I'm not using any plug-ins.

 

Frank

 

www.mp3.com/spigots

[miroslav]

OK...lessening the latency is the big questio here.

 

But I now have a secondary question.

Until latency is eliminatd or lessend to an "acceptable"(?) degree...how ARE some of you dealing with it in a real life situation such as described by spigots(Frank)?

 

So the musicians are bitchin' at you...what do you do to correct it so that they CAN play without that damn delay?

[Anderton]

Just confirming Al Pickard's post: most A/D converters I've measured spec out in the 600 - 700 microsecond range.

[GY]

Audio latency is a major issue when overdubbing music tracks, hard effects for a film or video project. It's like trying to win a 10k run with a broken leg. Is a ready made DAW like Protools the answer? Not for me. I need software with great MIDI/audio capabilities. I want it to reside on a machine that will run all my other programs. I want 128 midi tracks. I want at least 16 audio tracks. I want it all!

 

Sonar is promising, but missing the mark with me, so far (except for it's MIDI). I will keep reading and looking for new drivers to bring latency and reliability to a useable level. All the tweaks I've tried so far don't help.

 

I use 24 tracks of DA-88. I love it, but can't edit like I'd like. When I turn the system on, I feel confident it's going to allow me to finish my project. It syncs flawlessly. There's no latency. It "sounds" good.

[dansouth]

I haven't noticed any latency problems with soft synths. However, latency has been a problem with overdubs. That's one reason why I prefer to track overdubs on an AW4416, rather than a computer. I have a MOTU 828, now, and I may give that a shot since it features onboard monitoring.

 

Short delays may not be perceptible as delays, but they can cause undesireable phasing and comb filtering effects.