[Android] more then 20 people in the room, causese audio issue

[juunini]

Description

• When more then 20 people in the room
  • Android device white noise sound.
• When more then 25 people in the room
  • Android device no sound.
  • video, audio and people's talking everything. all sound.

To Reproduce Steps to reproduce the behavior:

1. More then 20 people in the room
2. Listen android device sound
   • my device is Galaxy S8, S10, Fold3
3. More then 25 people in the room (or 30)
4. No sound to listen in android device

Expected behavior

• Can listen white noise (more 20)
• Can't listen any sound (more 25 or 30)

Screenshots no screen shot. sound issue.

Hardware

• Device: Galaxy S8, Galaxy S10, Galaxy Fold
• OS: Android
• Browser: Chrome

┆Issue is synchronized with this Jira Task

[Dayk0]

Hello,

It's normal. Basically, hubs allows the possibility of having a maximum of 25 people in a single room. Depending on the device used (example: android phone) but also on the 3d environment in which you are, this can cause latencies. For there to be 25 people without sound problems, the environment must be the most optimized with simple avatars but also a good connection for each user.

[takahirox]

Hi, thanks for the problem report.

Yes, the maximum capacity of a room is 25. Maybe no audio quality guarantee for over 25 people.

https://hubs.mozilla.com/docs/hubs-room-settings.html#maximum-capacity

But 20 is under that maximum capacity. Probably worth to dig into the problem especially if the problem can even happen with fewer people in a room.

So let me ask some questions to investigate the problem more deeply.

1. If you access the same room from other non-Android devices, can the same audio problem be reproducible? I want to know whether the problem is Android specific or not.
2. Can the problem be mitigated if you use more optimized scene?
3. You seem to have several Android devices. Can the problem happen more easily on older (or lower-end) devices?
4. Can the autio graduately worse depending on the number of people in a room? Or does it suddenly happen once the number of people exceeds a certain threshold like 20?

[juunini]

@takahirox

1. Only android issue as far as I know. I'm not hear the white noise in non-Android device.
2. if Scene is more heavy, white noise early to hear when people exist 20. but, Scene is very light, I was hear the white noise people 28, 25 is in room and 3 is lobby.
3. Yes, low performance or old phone is early and loudly hear the white noise
4. I think the condition is when the device performs heavy calculations. The more people there are, the louder the white noise is.

[takahirox]

@juunini Hi, thanks for the answers. Do you mind if checking the "Disable audio left/right panning" option checkbox in the preferences and checking if the problem can be mitigated? You can find the option in "More" - "Preferences" - "Audio" tab.

[takahirox]

I tested and confirmed that the audio can be very choppy or even I hear nothing if people are many (about ten or more) in a room on my Samsung Galaxy S8+ + Chrome.

I realized that the problem can be worse if I move or rotate. So I speculate the problem can be related to Positional Audio (left/right panner), especially if CPU (or memory?) pressure is high. Actually if I check "Disable audio left/right panning" the audio problem is very mitigated.

If my speculation is true, potential solutions or workarounds may be

1. Force to disable Positional Audio on Android device or encourage to disable it if users face the audio problem
2. Figure out and optimize the performance bottlenecks in Hubs Client because the performance poorness seems to make the audio problem worse
3. Optimize the default scenes and avatars and encourage to use them for mobile devices
4. The root cause may be on Android end so make micro test which can reproduce the problem and report to Android devs

---

Can I ask for volunteer, android device users? Would you try the following steps and check if you can reproduce on your device? My device is Samsung Galaxy S8+ and it supports up to Android 9, doesn't support newer ones. So I feel there may be a chance that this Positional Audio problem can happen only on old Android.

1. Enter a room (I used the standard Gathering Hall for the scene) from your android device
2. Run ten or more bots with the following command on desktop

# replace roomId and roomName with your room's ones
$ ./run-bot.js -u https://hubs.mozilla.com/roomId/roomName -a ./bot-recording.mp3 -d ./bot-recording.json

3. Confirm that audio is very choppy on your Android device and the audio problem can be worse if you move or rotate
4. Check "Disable audio left/right panning" option in the preferences and the problem is mitigated

In my Samsung Galaxy S8+ Chrome the audio problem happened with about ten or more bots in a room. I entered the same room from Windows 10 Chrome and iPhoneX Safari and didn't see the problem on them. (I guess the reason why the problem didn't happen on my iPhoneX might be that we force to disable the Positional audio on iOS?)

[rawnsley]

I did try and run the bot script, but it appears puppeteer isn't working currently on M1 Macs :-(

[keianhzo]

I have been looking into this today using Galaxy Samsung S8 with Chrome. These were my tests:

• Empty scene Just two avatars, average CPU usage is 50%
• Avatars Adding over 4 bots makes the audio quality to degrade, over 8 it just stops. CPU usage goes up also and around 12 bots it's hitting 100% spikes. Removing the avatars one by one restores the audio quality.
• Positional Audios (panner nodes) Over 4 positional audios the quality starts degrading and I can appreciate small random cuts. Having 6 positional audios makes the audio stop. Removing them one by one restores the audio quality. The CPU usage after 6 positional audios shows spikes of around 100%. Removing the audios one by one restores the audio quality.
• Stereo Audios Over 10 stereo audios the quality starts degrading and I can appreciate small cuts. Over 20 the audio just stops and the CPU is around 90%. Removing the audios one by one restores the audio quality.

After this test I think it may not be related to the amount of avatars in the room (so not related to the WebRTC stack) and it's probably more related to the Panner Node performance in mobile devices or WebAudio pipeline issues.

Things that I'd try:

1. The Chrome EAC hack is not enabled in mobiles but I'd make sure that we aren't introducing any further complexity by mistake.
2. The output devices hack is also not enabled in mobiles although but same as before I'd check if that's not complicating the WebAudio routing unnecessarily.
3. WebAudio routing complexity. We have quite a few nodes and mixers going on for Panner Nodes, I don't think that it justifies the performance issues but maybe simplifying it to the minimum might help to uncover some issue.
4. Make a simple test that just creates panner nodes over times to see what's the actual Panner Node limit per hardware without all the Hubs client overload and compare to FF mobile. Maybe it's a Chrome mobile bug or limitation.

[takahirox]

@rawnsley Oh

@keianhzo

Thanks for the trial and comments. It sounds like we reproduced the same audio problem. What I want to know first is if this likely Positional Audio problem can be reproducible on newer Android devices at least which support the latest Android version, to determine how much we should take effort for investigating the problem. So waiting for volunteers who have newer devices.

And the problem seems to be very related to the performance poorness of maybe both (Positional) Audio related and unrelated so I'm thinking of finding the Audio related and non-related performance bottlenecks and trying to optimize them.

[netpro2k]

Tried this on my pixel 6 pro on Android 12

In Firefox I start getting noticeable crackling at around 4 bots, much worse at like 7, very bad > 10 In Chrome some minor crackling around 10, starts to get pretty bad at 13, rapidly gets worse after that

Some of what I am hearing just sound like the audio decoder is struggling to keep up, so I wonder how valid this testcase is. Its pretty rare you will have so many people actively speaking at the same time. Maybe a mix of speaking/non-speaking bots might be a good test case.

It might also be worth looking into adjusting the panning model on mobile. Three's PositionalAudio sets it to "HRTF", but I believe that is a good deal more costly. https://developer.mozilla.org/en-US/docs/Web/API/PannerNode/panningModel

Related to this, we looked into google's resonance audio SDK a very long time ago, it supposedly scales better with many audio sources https://resonance-audio.github.io/resonance-audio/develop/web/developer-guide though it would seem we should be under the number of sources where the performance breaks out.

[netpro2k]

With panning disabled: in Chrome I get up to around 11 without any issue but then it very quickly broke down after that. Not so much crackling but lots of audio skipping In Firefox I start to get some minor crackling around 8, gets pretty bad around 10 or 11, and then quickly breaks down after that

So it was slightly better without panning in Chrome, but more noticeably in Firefox. The full breakdown point in Chrome seems to be about the same, and sounds less like PannerNode issue and more like decoding issues like I described above.

(note its also pretty hard to tell if things are working right just because its impossible to actually listen to that many people talking at once and make sense of it hehe)

[juunini]

Galaxy Z Fold3 in same issue.

If scene is museum Not many people say. But many video say. T.T

[takahirox]

@netpro2k

Thanks for testing. I want to raise the priority of this issue because it's reproducible with about 10 people in a room even on Android 12.

Maybe a mix of speaking/non-speaking bots might be a good test case.

As we discussed internally, yes most codecs seems to have a concept of empty frames and WebRTC also has a way to flag a silent stream and doesn't even send data. Probably the less mics are on or less people speak in a room, the less we have audio decode pressure.

But note that this problem can even happen in practical use cases. I joined a few rooms where there were 10-15 people on my Android, and the audio was very cracky or even I couldn't hear anything. (There might be a chance that my old Android 9 can cause the proplem much easily.)

It might also be worth looking into adjusting the panning model on mobile. Three's PositionalAudio sets it to "HRTF", but I believe that is a good deal more costly. https://developer.mozilla.org/en-US/docs/Web/API/PannerNode/panningModel

Cool, it should be worth to dig into the positional audio parameters.

So it was slightly better without panning in Chrome, but more noticeably in Firefox.

Good to know that disabling the panner node can mitigate the problem. But bad to know that it isn't a perfect solution.

As @brianpeiris mentioned internally, iOS likely had the same problem and fixed, it won't be surprising if Android has the similar problems.

• https://bugs.webkit.org/show_bug.cgi?id=227199
• https://bugs.webkit.org/show_bug.cgi?id=234979

So in middle or long term we should (make micro tests which can reproduce the problem and) report to the Android devs. (I hope) the fact that iOS likely fixed the problem can push them.

Let me keep working on this problem.

Update: As we discussed internally https://github.com/gfodor/hubs-mpl/blob/master/src/systems/avatar-audio-track-system.js can help

[takahirox]

I made a micro test which can reproduce the problem.

https://takahirox.github.io/pannernode-test/index.html

Can I ask for volunteers again, Android 12 users? Would you please try the test and check if the audio can be broken on your devices?

On my Samsung Galaxy S8+ Android 9 Chrome, the audio is crackled with 14 or more panners with HRTF panning model, and no audio with 30 or more panners.

I want to report the problem to the Android devs once we confirm the problem can be reproducible in the test on Android 12.

Some notes:

• Using equalpower panning model can metigate the problem, but not a perfect solution
• The problem can be much mitigated if the panners and listener don't move
• Interesting thing. If I add some main thread CPU stress, the problem can be mitigated
• I couldn't reproduce the problem on my Windows 10 Chrome and iPhoneX Safari. Probably this problem is Android specific

[netpro2k]

All of these tests are on my Pixel 6 pro running Android 12:

Chrome: With HRTF, moving nodes, no added CPU load I can get up to about 30 without much crackling 30-50 crackling starts to ramp up, worse when interacting with the page Above 50 crackling continues to get work, and audio dies when interacting with the page Above 80, no audio

Chrome: With Equal Power, moving nodes, no added CPU load I can get up to about 40 or 50 without too much crackling, continues to ramp up fairly linear for the checkpoints I tested Audio plays all the way up to 100 nodes though with a lot of crackling and cutting out, interacting with the page completely cuts out the audio for several seconds, but it comes back.

Chrome: With HRTF, non-moving, no added CPU load I can get up to about 40 without too much crackling Audio plays all the way up to 100 nodes with a decent amount of crackling but no dropouts., interacting with page does not break things.

Chrome: With Equal Power, non-moving, no added CPU load I can get up to about 40 without too much crackling Audio plays all the way up to 100 without the crackling ever getting too bad (hard to tell at the high end since so much audio is playing), interacting with the page does not break things.

---

I didn't run all the tests in Firefox because things start to fall over pretty quick. Let me know if running more tests in Firefox would be helpful Firefox: With HRTF, moving nodes, no added CPU load At around 5 I already start to get some crackling and distortion getting bad pretty quickly above that I can play up to 100 but there is massive crackling and distortion. The audio also appears to be playing back at a lower pitch and rate. Interacting with the page does not seem to effect things.

Firefox: With Equal Power, non-moving, no added CPU load Craclking starts to get pretty bad around 30 or 40 At 100 pretty massive distortion and some crackling, but not the same slowed playback rate as above.

[takahirox]

Filed an issue https://bugs.chromium.org/p/chromium/issues/detail?id=1308962

Perhaps it takes at least some months to resolve the problem. I want to think of our short-term workarounds tomorrow.

[takahirox]

Thanks for testing, all.

As a short term workaround I think disabling panner node by default or forcing to disable it on Android device is good because it gives huge improvement and the change should be small.

Actually I entered a room having 18 people today on my Samsung Galaxy S8+. I couldn't hear with panner node enabled. By disabling it audio became fine.

We can think of other workarounds later if disabling penner node isn't good enough.

Switching panner model is interesting to me. But I want to think of it separately from the Android audio problem but as optimization.

[arpu]

Hi, old hubs/jel use this positional updateMatrixWorld hack: https://github.com/jel-app/jel-mpl/blob/master/src/hubs/utils/threejs-positional-audio-updatematrixworld.js

maybe this could be updated in more stable way

[arpu]

do you know, not using linearRampToValueAtTime in the matrixUpdateWorld solves the Android Problem?

https://github.com/mrdoob/three.js/blob/dev/src/audio/PositionalAudio.js#L116

[arpu]

have done some testing on android 12 pixel https://github.com/arpu/pannernode-test/commit/11ce03a59050a32f462a5aa583f6a79ef6d4918d https://arpu.github.io/pannernode-test/

with Equal Power, moving nodes, no added CPU load i get 60 panners running fine on 60fps

[vincentfretin]

With my S8 Android 9 Chrome 100, with https://takahirox.github.io/pannernode-test/index.html or https://arpu.github.io/pannernode-test/ I have the same results, with move panners and listener checked, none additional main thred CPU stress. With HRTF I start having audio issues around 16 panners. With equalpower, I can have some additional panners before I start having audio issues, around 23 panners.

[guanghw]

Can we fix this by limiting the maximum number of sound sources a single user can receive？ For example, it is calculated based on distance and sound volume, and only the 5 loudest sound sources are received.

[takahirox]

We have introduced the Audio PanningQuality preference in #5454 Try low quality mode. It isn't a perfect solution but may metigate the problem.

[takahirox]

I made a PR to use low audio panning quality mode by default on Android #5540 I hope we go with it and will see if the problem can still happen in practical use cases.

[takahirox]

5540 has been merged. Let us know if you still face the audio problem on Android devices in practical use cases.