Brain Neuroinformatics: Auditory System

1
Brain NeuroinformaticsAuditory System

• February 27, 2002
• Soo-Young Lee
• Brain Science Research Center
• Korea Advanced Institute of Science Technology
• http//braintech.kaist.ac.kr bsrc.kaist.ac.kr

2
Artificial Auditory System

• Based on Human Cognitive Mechanism
• Develop mathematical model and auditory chip
• Develop Continuous speech recognition system

3
Auditory Pathway

• ? Improvement of Cochlea model
• Binaural hearing model
• Auditory cortex model
• Selective attention model
• Continuous-speech recognition system
• Speech-recognition chip

4
 
Research Scopes
Auditory Pathway
 
5
Masking

• Lateral Inhibition
• Recursive

6
Masking using Lateral Inhibition
7
Temporal Masking
8
Temporal Masking Integration Model
9
Time Frequency Response
10
Isolated Word Recognition Rates
11
Discussions Masking

• Masking Suppresses Unwanted Noisy Components in
  Signal
• Simultaneous Masking by Lateral Inhibition
• - Recognition performance was enhanced with
  MFCC model
• - Proposed model can be used with any
  auditory model
• - ZCPA has spectral masking effect
• Temporal Masking by Unilateral Inhibition
• - Unilateral inhibition using the integration
  model
• - Model resembles other feature processing
  algorithms
• - Recognition performance was enhanced with
  RASTA parameters

12
Channel Compensation

• ASR in mismatched environments
• Environmental information
• Background noise, acoustic/transmission channel
• Assume environment degradation model

13
Speaker-to-Microphone mapping

• Mapper train
• Where and which type of mapper should be
  deployed?

14
Adaptive Noise Cancelling

• Adaptive noise cancelling
• An approach to reduce noise based on reference
  noise signals
• System output
• The LMS algorithm

15
ICA-based Approach to ANC

• The difference between the LMS algorithm and the
  ICA-based approach
• Existence of the score function
• The LMS algorithm
• Decorrelate output signal from the reference
  input
• The ICA-based approach
• Make output signal independent of the reference
  input
• Independence
• Involve higher-order statistics including
  correlation
• The ICA-based approach
• Remove the noise components using higher-order
  statistics and correlation

16
TDAF approach to ANC

• Normalized LMS algorithm
• Normalized ICA-based algorithm

where
where
17
Experimental Results (1)

• Experiments for artificially generated i.i.d.
  signals
• SNRs of output signals for the simple simulation
  mixing filter (dB)

18
Experimental Results (2)

• Experiments for recorded signals
• Signal waveforms for the car noise and the simple
  simulation filter

Signal source
Noise source
Primary input signal
System output signal
19
Experimental Results (3)

• Experiments for recorded signals
• SNRs of output signals for the measured filter

20
Experimental Results (4)

• Comparison of learning curves with and without
  TDAF
• Car noise

The ICA-based approach
The LMS algorithm
21
Discussion ANC

• A method to ANC based on ICA was proposed.
• The ICA-based learning rule was derived.
• The ICA-based approach
• Include higher-order statistics
• Make the output independent of the reference
  input
• The LMS algorithm
• Make the output uncorrelated to the reference
  input
• Gave better performances than the LMS algorithm
• TDAF method was applied to the ICA-based
  approach.
• Derived the normalized ICA-based learning rule
• Improved convergence rates

22
Bottom-Up and Top-Down Attention
External Cue
Internal Cue
Attended Output
Classifier Output

• Bottom-Up
• - Masking
• - ICA
• Top-Down
• - MLP
• - HMM

Bottom-Up Recognition
Top-Down Expectation
Attended Input
Bottom-Up Attention
Input Features
Brain
Environment
Input Stimulus
23
??? ?? ?? ????(1)

• ??? ?? ?? ????
• ??8051???? ??(???)
• 12 MHz ?? 1.93 MIPS
• ???? ?? ??(16??)
• ?? ??? ?? ?? ???
• 50 ?? ? 20 ms ?? ?? ?? ??
• ??? 95 ??

24
??? ?? ?? ????(2)

• Hynix 0.35 ??
• ??? ?? 64-TQFP
• ?? ??? 128x12, 256x8, 2048x8
• AGC, A/D, D/A ??
• 12KHz, 12 bit A/D
• 12Khz, 8 bit D/A
• ??? ?? ?? ????
• ?? ?? ????
• ?? ?? ????
• ??? ?? ???? ??

25
??? ?? ?? ????(3)

• CM8051 ??? ? I/O ???

26
??? ?? ?? ????(4)

• ?? ?? ???

27
Analogue ICA Chip

• ? ? multiplier
• ? ? current summation
• r ? learning rate

4 x 4 ICA network in one Chip
28
Test Results in Waveforms

• Source Signals (s1, s2)
• two different males voice
• 16 kHz sampled
• Mixed Signals (x1, x2)
• Instantaneous mixture
• Mixing Mtx A is
• Separated Signals (o1, o2)
• Recovered original sources

29
Fabricated Chip

• 2.8mm x 2.8mm
• AMS CMOS 0.6um
• 2 poly-3 metal
• Analog Digital Hybrid Process

Die Photo of a Fabricated ICA Chip