Latency

[midimusic]

The subject is Latency.

What is the scientific definition of latency with regard to organs? And how is it to be properly measured? If latency cannot be measured then latency figures quoted for organs are almost meaningless.

In simplistic terms latency is obviously the time between pressing a key and the emission of sound from a "pipe". But this in itself has no scientific meaning. How is the time of pressing a key measured? How is the time a sound is emitted measured? Both must be defined before latency can be properly measured. I am not aware of any scientific definition of either.

When you press an organ key, at some point in its depression it opens a valve to pass air to a pipe. All this takes time so at what point is the key defined as being pressed?

When an organ pipe begins sounding the sound starts from nothing and gradually builds up to a steady state. At what point is the sound defined as being emitted (for the purpose of latency measurement). This can make an enormous difference to a latency measurement.

I am not expecting this to be resolved here but would be interested in other peoples thoughts. Especially whether they believe latency figures quoted for organs have much meaning.

csw900

[newVideophile]

A while ago, I was working on a fast piece at full speed, and I noticed that I was making a lot more mistakes than usual. After wondering why I was suddenly so bad, I had the presence of mind to look at the latency in GO. The value had risen to 40 ms, because (for some reason) the settings had changed. I reset my original configuration, with latency displayed at 20 ms, and started playing correctly again!

So, whatever GO's way of measuring latency, the value it displays makes sense.

Personally, I see 4 steps:

1. The organist decides to press the key
2. From a certain position, the key "understands" that it is depressed
3. The command then travels (mechanically/electrically/electronically, etc.) inside the organ until it triggers the opening of a valve
4. The sound unfolds, with a different build-up, of course, depending on the note and stop used, until it becomes perceptible.

For me, total latency is therefore the time between the start of stage 1 and the end of stage 4.

The latency of step 1 takes place inside the organist and can therefore be removed from the equation.

For step 2, the question is essentially to know at what level of depression the key will detect pressure. I suppose this varies from organ to organ. On mine, I've made a 60/40 setting: start playing if key >= 60%, stop if <=40%.

Step 3 poses no problem of definition. In fact, it's the one that best defines the organ's intrinsic performance.

Step 4 is the most problematic, because as you mention, the criterion chosen (i.e. the sound level threshold) can have a considerable influence on the result. Besides, in real life, the perceived latency of a note will not be the same if played alone or with other notes at the same time.

My two cents!

[willeke1234]

Is there a step 5? The sound goes from the pipe to the organist.

[midimusic]

Yes the time the sound takes to travel from the pipe to the organist can be significant in a large room such as a church.
csw900

[midimusic]

Well it looks as if either nobody else is reading this or they have nothing to say about latency.

newVideophile has expanded my original comment which was intended to be completely general and scientific - hopefully moving towards a proper scientific measurement of latency.

First the mention of figures for latency - apparantly GO "displays" latency - where is this display and is it just a guess or a real measurement? I have not found it. Do you really believe the figures 20ms and 40ms? I have doubts..

If we are now talking only about GO, "Step 3" is the only area where GO could attempt to measure its own latency. It would need to start a high resolution timer when it received a "note on" midi message from rtmidi and stop it when it told the operating system's sound driver to play the note. It would then need to display the result. This would all be possible but as far as I know it does not do this.

I remember the days before computers had hardware sound systems. Latency could easily be 500ms - making playing music from a keyboard next to impossible. Nearly all of the improvement since then has been due to the improvements in computer hardware not in VPOs (software organs).

csw900

[willeke1234]

According to the help "Audio/Sound Output State tries to displays actual latency.". Choose Audio/Sound menu -> Sound Output State. It's a calculation with samples per buffer and sample rate.

[midimusic]

I found this and it says 69ms on my windows system. It would be useful to know the exact formula GO uses to make this calculation in order to check its believability. On this windows system GO plays plays from a keyboard without ANY NOTICEABLE latency.

csw900

[oleg68]

@midimusic

A way of determining an actual latency might by a special device with a midi-out interface and a microphone that mesures the time interval between sending a key-on midi event and capturing the sound from the speaker. I don't know about existance of such devices.

We may build a software solution: a program that emulates midi messages and listen for an audio input. It eliminates a portion of latencies for midi encoding, converting midi to usb, receiving a usb and decoding midi, but it is usually quite small (unless we use usb over wifi/ip tunelling).

This testing program may listen not an audio input from a real microphone, but a virtual audio input connected with a GrandOrgue audio output. In this case we don't take into account a latency added by an audio device. Usually professional or semi professional audio cards add very small latency, but audiophile DACs or HDMI sound output with an external AV-receiver may add quite large latency that will not be measured in this case.

Usually the largest part of latency is brought by GrandOrgue and other sound processing software behind it, because they need to fill an audio buffer and pass it to the next step of audio peocessing chain. So increasing a sample rate, decreasing a buffer size and eliminating extra steps of the audio processing chain decrease the latency.

20ms and 40ms are not precise and they do not reflect the actual latency.

Usually a high latency does not bring any problems for playing midi files, but it becomes an obstacle for a live organ performation, that @newVideophile mentioned above.

Real pipe organs also bring some latency for playing. Especially it is true for some pneumatic-tracker ogans, ex. built by Walker.

Usually low notes of pedal divisions have long attacks: they require a significant time to grow the sound apmplitude after they started playing. You can see this by opening their attack wav-samples with a sound editor, ex. Audacity.

[midimusic]

Many years ago I tried this method of properly measuring latency. All you need is a microphone and sound recorder. Hit the key with something that makes a noise. The resulting recording is then displayed in an audio analyser and the times can be easily determined. However the rise time of the audio waveform is by far the longest time measured and there is no standard to say at what point the audio began. At the first instance the wave can be seen it will be too soft to hear - so at what point does the sound become audible? I would suggest that 70% of the final steady state would be reasonable.

This is a complex subject and as I said above I do not know of any standards.

csw900

[midimusic]

I have just done a brief test using GO demo and the results are:
Time to first visual evidence of the note (middle A) around 60 ms.
Time for note to become audible around 100ms.
Buffer size = 128 with 44100 sps stereo.

csw900

[midimusic]

Regarding the GO sound output buffer size. My previous experiences with this indicated that the optimum buffer size for all systems is 128. I am disappointed to note that the old default of 1024 (which does not work properly on any system; appears to be still in place and continues to fool others as well as me). This high value (1024) does compromise the maximum speed at which the organ can be played - including the playback of midi files.

I have just confirmed on my windows system the size of this buffer has no significant effect on latency - I can find no noticeable change in human detectable latency for any values betwee 128 and 1024. This is possibly confirmed by "Sound Output State" readings which for my system are as follows for sample rate 44100 (Note that my careful spacing of this table has been ruined by Github's ignoring white space characters!!).
Buffer size, Latency ms
4096, 69
2048, 69
1024, 69
512, 58
256, 58
128, 52
64, 50
32, 50
16, 50

As I said above it would be very useful to find out from the code how these values are calculated. At present I have doubts about the validity of all latency figures displayed by GO. Hence the desirability of a proper scientific measurement.

csw900

[midimusic]

@willeke1234

It is an interesting subject which nobody seems to have done before. There can be no meaningful discussion about latency if nobody has measured it. We now know that the very low figures often quoted have no useful meaning in reality.

It is a pity that a few more GO users have not joined in this discussion. Maybe it is because they are quite happy with 100ms latency.

csw900

[willeke1234]

@midimusic
I will do some measurements of the total latency if you tell me how. The recording looks like this:

How do I measure latency? I can see where to start, but how do I determine the end?

[midimusic]

Your waveform looks very much the same as those I get. You can see the peaked wave at the start - time begins at the first point you can definately see where this wave began. In this case it is at start of the low level wave before the signal begins to rise steeply. This is the time you hit the key and is time 0.

The silent latency period ends where you first see see the organ note beginning to rise. Measure this time. In this case I would say it is at the start of the downward sloping wave just before the silent crossover point where the organ wave starts rising in amplitude. Keep a record of this time.

The latency time is the total time to the the point where the amplitude of the organ note reaches 70% of its final steady state value. Keep a note of this time.

Save the file with a name that records accurately what it is. Also keep written notes if necessary. You should be able to repeat this measurement at a later date and get the same results.

I hope this is helpful.

csw900

[willeke1234]

I don't think I can measure latency reliably. The steady state value is not stable, probably because I'm recording with an iPad. And if it was stable, it is a challenge to determine the 70% point.

[midimusic]

@willeke1234

Many organ pipes do not produce a stable steady state value. However you do ideally need something better than an iPad to make accurate measurements. I used an old iMac with windows 11 and Adobe Audition to make my masurements and still I would not say that the 70% level is easy to determine but at least an approximate value can be found.

When there are no measurements available the best you can do is always better than nothing. A scientist always puts a +-nn% tolerance on his measurements to indicate how accurate they are.

I think you have given up too easily - please try again!! Alternatively if you attach a few waveform files me or someone else can analyse them for you.

csw900

[willeke1234]

I think the input level is adjusted automatically, so the levels are unreliable. With a buffer size of 128, my latency is 90-100 ms. With a buffer size of 1024 it is 140-160 ms.

[manu13320]

I think latency is difficult to evaluate by ear with organ sounds; a more nervous instrument is more suitable, such as the harpsichord, for example http://duphly.free.fr/en/blanchet.html
An acceptable latency would be 50 ms, 100 ms already seems too much to me.

[midimusic]

@manu13320
I use the blanchet harpsichord sampleset. It sounds OK to me.
I invite you to MEASURE its real latency rather than quoting almost random figures. A harpsichord's shorter rise time will automatically make the measured latency lower. But it will still depend mainly upon the organ it is being played on.
csw900