1. EPSRC Perceptual Constancy Meeting
Wednesday 19th May

2. Sheffield Overview Guy J. Brown and Amy Beeston
Department of Computer Science
University of Sheffield

3. Schedule Brief overview of work to date (Guy)
Efferent model with within-channel compensation mechanism (Amy)
Perceptual experiments on the AI corpus (Amy)
Automatic speech recognition/modelling with the AI corpus (Guy)

4. Brief overview of work to date

5. Perceptual compensation experiment (Watkins, 2005)
Test word drawn from continuum between “sir” and “stir”
Carrier : near
Test : near
Tendency to stir
Carrier : near
Test : far
Tendency to sir
Carrier : far
Test : far
Tendency to stir  Compensation
OK, next you’ll get {TEST} to click on
Note: “near” and “far” indicate source/receiver distances of 0.32m and 10m respectively

6. Does compensation arise from dynamic range control?
Does the auditory system identify and compensate for reverberation tails?
Alternative proposal: might compensation in Watkin’s experiment arise as a consequence of mechanisms that control dynamic range?
Reverberation reduces dynamic range
Illustrate using blurredness metric: mean-to-peak ratio (MPR)
Near-near condition
Distance 0.32m
MPR = 0.19
Far-far condition
Distance 10m
MPR = 0.32

7. Possible role of the efferent system?
Compensation involves restoration of dynamic range
Efferent system (MOC) implicated in control of dynamic range via closed-loop feedback (Guinan & Gifford, 1988)
Plausible time scales
Efferent feedback is sluggish, in the range ms. Also long term effects over tens of seconds (Sridhar et al. 1997).
Contrast adaptation at higher levels (IC, cortex)
Dean, Harper and McAlpine (2005)
Neil Rabinowitz (Oxford)
Neurons shift their level-response functions to maximise coding efficiency relative to the distribution of levels

8. Template-based recogniser
Schematic of the model
DRNL
Outer & Middle Ear
Hair cell
Framing
Template-based recogniser
AN response
Metric (MPR)
Efferent attenuation
Auditory periphery
Efferent system
Stimulus

9. Dual-resonance nonlinear (DRNL) filter
stapes velocity (m/s)
3 Butterworth filters
3 Gammatone filters
Broken stick
nonlinearity
‘Efferent
attenuation’
MOC (ATT)
4 Butterworth filters
2 Gammatone filters
Linear gain
BM velocity (m/s)
Nonlinear path
Linear path
Proposed by Meddis, O’Mard and Lopez-Poveda (2001), human parameters from Meddis (2006)
Efferent attenuation introduced by Ferry and Meddis (2007)

10. Effect of efferent suppression
Hair cell is simple threshold and rate limiter as described by Messing (2007)
Efferent suppression causes rate-level function to shift to the right
Low-level activity (on toe of the curve) falls to spontaneous rate

11. STEP: effect of efferent suppression

12. Feedback control of efferent attenuation
Closed loop system obtained by computing MPR of pooled AN response within a 1 s sliding window
Efferent attenuation updated every 1 ms
Efferent attenuation increases linearly with MPR, tuned according to near/near and far/far conditions
MPR
ATT = f(MPR)

13. Results
Human listeners (Watkins, 2005)
Computer model

14. Analysis - why does it work?
Higher MPR, more efferent suppression when context reverberated at far distance
MPR is changed very little by time-reversing the speech (dynamic range largely unaffected)
MPR is reduced in the region preceding the test word when the reverberation is reversed.

15. Summary Good model of some of Tony’s data.
Plausible model in terms of auditory physiology
Still lots to do:
noise contexts
within-channel compensation mechanisms (Amy)
modelling vocoded stimuli (in progress, Amy and Simon)
model of a specific experimental finding or more general processor that can be applied to ASR?