Acoustic Localization by Interaural Level Difference

1
Acoustic Localization by Interaural Level
Difference

• Rajitha Gangishetty

2
Acoustic Localization
Acoustic Localization Determining the location
of a sound source by comparing the signals
received by an array of microphones. Issues
reverberation noise

3
Overview

• What is Interaural Level Difference (ILD)?
• ILD Formulation
• ILD Localization
• Simulation Results
• Conclusion and Future Work

4
Techniques

• Interaural time difference (ITD) relative time
  shift

• Interaural level difference (ILD) relative
  energy level

All previous methods (TDE, beamforming, etc.) use
ITD alone.
5
Previous Work

• Time Delay EstimationM. S. Brandstein, H. F.
  Silverman, ICASSP 1997 P. Svaizer, M.
  Matassoni, M. Omologo, ICASSP 1997
• BeamformingJ. L. Flanagan, J.D. Johnston, R.
  Zahn, JASA 1985R. Duraiswami, D. Zotkin,
  L.Davis, ICASSP 2001
• Accumulated CorrelationStanley T. Birchfield,
  EUSIPCO 2004
• Microphone arrays Michael S. Brandstein,
  Harvey F. Silverman, ICASSP 1995 P. Svaizer, M.
  Matassoni, M. Omologo, ICASSP 1997
• Hilbert Envelope ApproachDavid R. Fischell,
  Cecil H. Coker, ICASSP 1984

6
A sneak peek at the results
Likelihood plots, Estimation error, Comparison of
different approaches
7
ILD Formulation
8
ILD Formulation
9
ILD Formulation
10
Isocontours for 10log(delta E)
11
ILD Localization
Why multiple microphone pairs?

• With only two microphones source is constrained
  to lie on a curve
• The microphones cannot pinpoint the sound source
  location
• We use multiple microphone pairs
• The intersection of the curves yield the sound
  source location

12
Combined Likelihood Approach
Localize sound source by computing likelihood at
a number of candidate locations

• Define the energy ratio as

• Then the estimate for the energy ratio at
  candidate location is

where is the location of the ith
microphone

• is treated as a Gaussian
• random variable

• Joint probability from multiple microphone
• pairs is computed by combining the
• individual log likelihoods

13
Hilbert Transform

• The Hilbert transform returns a complex sequence,
  from a real data sequence.
• The complex signal x xr ixi has a real part,
  xr, which is the original data, and an imaginary
  part, xi, which contains the Hilbert transform.
• The imaginary part is a version of the original
  real sequence with a 90 phase shift.
• Sines are therefore transformed to cosines and
  vice versa.

14
Hilbert Transformer
In Frequency domain, Xi(ejw) H(ejw)Xr(ejw)
The Hilbert transformed series has the same
amplitude and frequency content as the original
real data and includes phase information that
depends on the phase of the original data.
15
Hilbert Envelope Approach

• All-pass filter circuit produces two signals with
  equal amplitude but 90 degrees out of phase.
• Square root of the sum of squares is taken.

16
Simulated Room
17
Simulation Results

• The algorithm
• Accurately estimates the angle to the sound
  source in some scenarios
• Exhibits bias toward far locations (unable
  to reliably estimate the distance to the
  sound source)
• Is sensitive to noise and reverberation

18
Results of delta E Estimation

• The estimation is highly dependent upon the
• sound source location
• amount of reverberation
• amount of noise
• size of the room
• relative positions of source and microphones

19
Likelihood plots
20
Likelihood plots
21
Likelihood plots
22
Likelihood plots
23
Likelihood plots
24
Angle Errors in a 5x5 m room
25
Angle error in degrees for the 5x5 m room when
the source is at a distance of 1m
Angle error in degrees for the 10x10 m room when
the source is at a distance of 1m
26
Angle error in degrees for the 5x5 m room when
the source is at a distance of 2m
Angle error in degrees for the 10x10 m room when
the source is at a distance of 2m
27
Comparison of errors with the Hilbert Envelope
Approach in a 5x5 m room
28
Comparison of errors with the Hilbert Envelope
Approach in a 10x10 m room
20 dB
10 dB
0 dB
Without Hilbert solid line, blue Matlab
Hilbert dotted line, red Kaiser
Hilbert dashed line, green
0.7
Reflection coefficient
0.8
0.9
29
Likelihood plots without Hilbert Envelope
30
Frames approach
5x5 m room, theta 18 deg , SNR 0db,
reflection coefficient 9, d 2m (left), d 1m
(right)
Mean error 15 deg Std Dev 11 deg
Mean error 7 deg Std Dev 6 deg

• Signal divided into 50 frames
• Frame size 92.8ms
• 50 overlap in each frame

31
Conclusion and Future Work

• ILD is an important cue for acoustic localization
  
• Preliminary results indicate potential for ILD
  (Algorithm yields accurate results for several
  configurations, even with noise and
  reverberation)
• Future work
• Investigate issues (e.g., bias toward distant
  locations, sensitivity to reverberation)
• Experiment in real environments
• Investigate ILDs in the case of occlusion
• Combine with ITD to yield more robust results

32
Thank You