Audio spatializer testing umbrella bug

[mcclure]

Premise of this bug:

1. We need a good, useful test or set of tests for the audio spatializers. It should test both location and reverb/geometry.
2. We need to run this test on all spatializers and conclude the results subjectively "sound good"

My proposal for (1) is that I'll port the spatializer test from my game project/fork to run on vanilla Lovr with ogg-file sources.

Re: (2), here are the outstanding known "sounds bad" bugs:

• Steam spatializer cuts off reverb when the source reaches its natural end Phonon reverb can get cut off #404
• Steam spatializer applies reverb even to non-spatialized sources Steam spatializer applies reverb even to non-spatialized sources #425
• Oculus reverb does not work Oculus spatializer reverb effect support (attempt) #422

Outstanding "functionality"/non-sound-related spatializer bugs:

• setGeometry is slow and can cause visual hitches / AudioMesh would be nice lovr.audio.setGeometry with phonon is kinda slow #406
• Box model option would be nice (Audio Geometry Box Model #408 / Oculus spatializer reverb effect support (attempt) #422)
• Should be possible to set materials for just part of a mesh (Per-triangle AudioMaterials #409)

The current best test we have

Check out the branch break-phonon-test-2 of my lovr-docs repo and run Showcase/Break. This version of break has some tweaks to support testing, it calls setGeometry, it enables all effects on the ball bounce sounds (but no effects on the "you lose" / "game starting" sounds), and the sounds have all been edited to have long periods of silence afterward (to work around #404).

You can select your spatializer/driver by uncommenting lines in conf.lua, and these lines at the top of main.lua let you quickly change the test:

local spaceSize = 16             -- Diameter of space 
local spaceFlip = false          -- If true reverse model winding
local audioMaterial = "concrete" -- Options include "carpet" "concrete" "glass"
local geometryMode = "mesh" -- Options "disable" "box" "mesh". "box" is oculus-audio-only
local volume = 1

There is a line you can edit later down to make the space a pentagon instead of a box.

Current test status

Testing with 0a4a62f from #422

• Steam audio selected in conf.lua:

  • HRTF sounds good.
  • I primarily tested with audioMaterial=glass. spaceSize=16 sounds "pretty good". spaceSize=8 sounds fakey/chintzy. spaceSize=160 sounds freakish, the echo is drawn out and sounds strangely robotic.
  • The echo is being applied to break-buzzer.ogg and break-countdown.ogg, even though those do not have any effects enabled.
  • GeometryMode "Disabled" works as expected

• Oculus audio selected in conf.lua:

  • HRTF sounds good.
  • Regardless of geometry node, no reverb is applied. Reversing winding has no effect.

[bjornbytes]

Finally starting to dig into this. This is a great summary of the current state of things and where they should improve. The Break branch is already a really good test case.

As mentioned in #425 I was able to fix the start/lose sounds to not have effects applied, I think this was just an issue with API usage.

One thing I noticed was really weird: If I uncomment the third setGeometry so it's applying the room mesh, on certain runs of Break, every time a blip plays this horrible distorted diminished chord blares like a train horn (which is really cool but probably a bug). Are you able to get that too? I was able to get it on all materials, intermittently. I think this means it's a race condition or something?

I'm going to initially focus on #404 so reverberating phonon sources don't need to be padded because that's just brutal UX

[bjornbytes]

If I create the blips with { effects = { 'reverb' } }, the train horn still appears, but if I do { effects = false }, the problem expectedly goes away.

[mcclure]

Oculus Audio now works in my testing (slightly updated both PR and test). In mesh mode it sounds bad because we can't set the material so it just picks a "perfectly reflective" material which isn't physically accurate.

I think I managed to trigger a bug with what happens if you run out of slots, I think instead of correctly killing old sounds it starts playing weird distortion.

[bjornbytes]

Do you know if it's when you run out of Oculus slots or Source slots? I did test the Source recycling stuff, Source:play should return false if you try to call play on it when too many (64) sources are already playing.

[mcclure]

It would be oculus slots. It was when the sounds were done playing but their tails were still running