SoundSource Localization using Interaural Time Difference via CrossCorrelation

1
Sound-Source Localization using Interaural Time
Difference via Cross-Correlation

• Dr. Harry Erwin
• harry.erwin_at_sunderland.ac.uk

Hybrid Intelligent Systems University of
Sunderland St. Peters Way, Sunderland, SR6 0DD
2
Acknowledgements

• John Christopher Murray, who did the work.
• Stefan Wermter, who helped me supervise the work.
• The staff of the Hybrid Intelligent Systems Group.

3
Aim of paper

• Show a system for robotic sound-source
  localisation
• System based on the mammalian auditory system
• Demonstrate a robotic system giving comparable
  results to that of the mammalian system

4
Service Robotics
Robotic system should be able to detect a
sound-source of interest and orient to face
that sound-source.

• Ultimately sounds will be detected with respect
  to background clutter.
• Tracking of the sound-source is of importance

5
Interaural Acoustic Cues

• To estimate Azimuth specific Cues are used
• Interaural Time Difference - ITD
• Interaural Phase Difference - IPD
• Interaural Level Difference - ILD
• t0- Sound reaches R
• t1- Sound reaches L
• Denoted by line a
• ITD t1 t0

6
Cross-Correlation to estimate ITD

• Used as a pattern matcher to check position of
  MAXIMUM similarity
• Independent sound signals g(t) h(t) are slid
  across each other (Sliding Window see diagram)
• Correlation vector is returned showing delay
  between the signals g(t) h(t) i.e. the ITD

7
Correlation Vector

• Maximum value in vector represents delay of
  signals g(t) h(t)
• Correlation vector delay is proportional to ITD
• In phase signals would produce MAX point at
  middle of correlation vector
• To determine delay locMAX (Vectorsize/2)
  Delay

8
Experiments

• Mounted two non-isotropic microphones on robot at
  a distance of 30cm from each other.
• 200ms sound recorded.
• Sound processed by cross-correlation
• Trigometric function used to compute final angle
  of incidence.
• Loop back to 1

9
Process
Trigometric Formula

• Pseudo Code
• azimuth()
• record_sound(buf, 200ms)
• normalise_data(buf)
• create_left_right_streams
• delay xcorr(left, right)
• trigometric_equations(delay)
• return angle

Where ? time between delay samples s value
of from cross-correlation C distance between
microphones
10
Results
Each angle over a repetition of five (5) times.
11
Efficiency

• Reaction time to compute angle
• 3 seconds on Lena (AMD K6-2-500MHz)
• 1 second on P4 2GHz Laptop
• Accuracy of sound localization
• Average - 1.5O
• Smaller errors at 0O
• Larger errors at 90O

12
Conclusions

• Human accuracy 0.9O 1.5O
• Our system 1.5O (Average)
• This shows a comparable system to that of the
  mammalian auditory system for sound source
  localization.
• Improved accuracy can be made with complex high
  sample rate computer sound cards.

13
Questions