Precomputed Wave Simulation for Real-Time Sound Propagation of Dynamic Sources in Complex Scenes

Nikunj Raghuvanshi†‡, John Snyder†, Ravish Mehra‡, Ming C. Lin‡, and Naga K. Govindaraju†

†Microsoft Research ‡University of North Carolina at Chapel Hill

Sound Propagation

• Essential for immersion• Physically complex, perceivable effects

Positive pressure Negative pressure

State-of-the-art: Games

Hand-tuned filters: difficult & tediousSimulated propagation → automatic, detailed, scene-dependent acoustics

Direct path muffledReflections muffled

Direct path muffledReflections clear

Direct path clearReflections muffled

Images: Dmitri Gait © 2003

Previous Work

• Beam-tracing [Funkhouser et. al. 2004]

• Frustum-tracing [Chandak et. al., 2008]

• Image-source method [Tsingos, 2009]

• Geometric methods: diffraction/scattering, high-order reflections challenging

Our Approach

• Wave simulation– diffraction: attenuation behind obstructions,

low-pass filtering– scattering from complex surfaces– dense reverberation

• Precompute & render– similar to [James et al. 06]– but propagation, not radiation

Comparison with Half Life 2TM

• “Train Station” scene

• Our engine vs. game’s default sound engine

Precompute & Render

Impulse ResponseTime

Pres

sure

Precompute

direct

reflected

Render

?

+=Pr

essu

re

Time*Convolution

Moving Source & Listener

Time

Pres

sure

Moving Source & Listener

Time

Pres

sure

Moving Source & Listener

Time

Pres

sure

Moving Source & Listener

Time

Pres

sure

Moving Source & Listener

Time

Pres

sure

Main Challenge

• Impulse Responses vary spatially• 7D sampling space

– source (3D) x listener (3D) x time (1D)

• Can reduce to 6D– Restrict listener to 2.5D surface

• Brute-force sampling still infeasible

Brute-force sampling

• Scene: “Citadel” (28m x 60m x 32m)• Simulation grid (12 cm)

– 200,000 gigabytes

• Sub-sampled (~1m)– 60 gigabytes– Interpolation non-trivial

Main Contributions

• Compact representation for impulse responses– 100x reduction in memory usage– works with (band-limited) wave simulations– plausible high-frequency extrapolation

• Spatial interpolation allowing coarse (~1m) grid• Real-time wave effects

– automatic correspondence to scene geometry

Our approach: Overview

Source locations

Listener locations

~ 1-2 meters

Wave Simulation

Response (Listener at source location)

Encode this compactly

• Simulator: ARD [Raghuvanshi et. al., 2009]• Fast and memory-efficient• Band-limited (~1kiloHertz) – time/memory constraints during precomputation

Source (Gaussian pulse)

Positive pressure Negative pressure

Psycho-acoustics

Time

Pre

ssur

e

Early Reflections

Late Reverberation

• Direct Sound: sense of direction• Early Reflections: loudness, timbre; spatial variation; 50-100 ms• Late Reverberation: decay envelope; no spatial variation; few seconds

Reference: “Room Acoustics” by Heinrich Kuttruff

Demo

• Perceptual effect– Only Direct sound– Direct + Early Reflections– Direct + Early Reflections + Late Reverberation

TimePres

sure

Record response

Probe Source (Gaussian)

Listener

Technique: Wave Simulation

Technique: Peak Detection

• search for local extrema• extract peak delays and amplitudes• wide-band information

Early Reflections (ER) Late Reverberation (LR)

Technique: ER/LR Decomposition

• separation based on echo density (500 peaks/sec)• single LR filter per room• 10x reduction in memory & precomputation time• ER varies spatially

Store one per room

Early Reflections (ER) Late Reverberation (LR)

Time Frequency

Technique: Frequency Trend Extraction

Band-limitedFFT

FFT

Technique: Frequency Trend Extraction

• peak data does not capture downward trend (diffraction)

Early Reflections (ER) Late Reverberation (LR)

Time Frequency

FFT

FFTDivide

Frequency trend

Extrapolated

Technique: Frequency Trend Extraction

Early Reflections (ER) Late Reverberation (LR)

Time Frequency

FFT

FFTDivide

Store ,[ ]Peak times and amplitudes Frequency trendExtrapolated

Technique: ER Representation

Early Reflections (ER) Late Reverberation (LR)

Time Frequency

FFT

impulse response interpolation

Runtime: Spatial Interpolation

listenersource

Runtime: Spatial Interpolation

Time

Pres

sure

P1

P2

P2

Time

Pres

sure

P1

Ours

Time

Pres

sure

Linear

Time

Pres

sure

peak aliasing

no aliasing

?P1 P2

Runtime: Render

• Fast frequency-domain convolutions– bottleneck: FFT– Performed using Intel MKL (single-threaded)

Results: Citadel Walkthrough

• Game environment from Half Life 2TM

• Size: 60m x 28m x 22m

Results: Walkway

• Size: 19m x 19m x 8m• Realistic attenuation behind walls

Results: Walkway

• Sound focusing – Loudness increases beneath concave reflector

Results: Living Room

• Empty vs. furnished living room

• Scattering off furnishings changes acoustics

Integration with Half Life 2TM

• Train Station scene: 36m x 83m x 32m

Cost

• Run-time: ~5% of a core per source• Memory: few hundred MBs• Precompute: few hours per scene

• Machine: 2.8GHz Quad-core Intel Xeon, 2.5GB RAM

Conclusion

• First real-time wave-based propagation engine– diffraction, occlusion/obstruction, late reverberation, focusing,

complex 3D scenes

• Moving sources and listener• Compact IR encoding – 100x compression • Fast, high-quality IR interpolation on coarse (~1m) grid

Future Work

• Need further 10-100x reduction in memory– 4x: 32-bit peak/frequency trend data → 8-bit on decibel scale– temporal IR compression: loudness, comb-filtering etc.

– spatial compression + adaptive sampling

• Scalability: cells and portals for sound– efficient boundary integral for portal transfers

• Limited dynamic geometry – Precomputed frequency-dependent occlusion factors

• Runtime: Use GPU for convolution (FFT and IFFT)

Thank you!!

• Go to: http://research.microsoft.com/en-us/um/people/nikunjr/siggraph2010/FastWave.avi

• Hear the video over headphones

Runtime Processing: RenderingPe

r Ear

.FFT

IFFTOutput

Short FFT

Pre-baked

Input

ER

LR

Frequency Trend. + +

Source 2

Source n

: element-wise multiplication.: element-wise sum+

Overview

• State of the art• Our approach• Results• Conclusion

Overview

• State of the art• Our approach• Results• Conclusion

Overview

• State of the art• Our approach• Results• Conclusion

Overview

• State of the art• Our approach• Results• Conclusion

Overview

• State of the art• Our approach• Results• Conclusion

State-of-the-art: Games

Image © Dmitry Gait 2003

Occlusion

Obstruction

Exclusion

Hand-tuned filters: tediousSimulated propagation → automatic, detailed scene-dependent acoustics