A physiologically motivated gammachirp auditory filterbank Toshio Irino
A physiologically motivated gammachirp auditory filterbank Toshio Irino (NTT Communication Sciences. Lab. Japan) Masashi Unoki (CNBH, Univ. Cambridge/JAIST) Sept. 2000 – August 2001 Roy D. Patterson (CNBH, Univ. Cambridge)
Early History Human Masking Data Roex filter Mag. Spectrum Gammatone AF Schofield and Cooke, 1985 Gammachirp filter Irino and Patterson, 1997 Well-defined impulse response
Recent Developments (1) Physiological Data Gammachirp de Boer and de Jongh (1978) de Boer and Nuttall (1997) Gamma tone chirp
Recent Developments (2) Chirp in Carney’s Revcor data (1999) » level-independent Chirp in BMM observed post-mortem at high SPL (Recio et al. 1998). » chirp is a property of passive BM response Compressive gain in membrane motion Disp. Compressive » also in human masking data » on frequency but not off frequency Input Level
Chirp in Revcor data (Carney et al. , 1999) Zero crossings Waveform Instant. Freq. Noise Level chirp Not level dependent
Where is level dependency? Analytic gammachirp (Irino and Patterson, 1997) GT Decomposition type GT Level dependent shape (Irino and Unoki, 1999) Asymmetric func. LP-AC HP-AC Physiological gammachirp (Irino and Patterson, GT LP-AC HP-AC Level dependent cf 2000) Level independent shape
Analytic gammachirp Gammachirp is a gammatone times an asymmetric function
Decompose asymmetric function Passive BM a fixed lowpass AF and a variable highpass AF
Physiological Gammachirp Fitted to human masking data of Rosen and Baker (1994), at 2 k. Hz Passive BM Gain change tails must converge tails converge Vary centre frequency of highpass, asymmetric function with level
Physiological gammachirp Fitted to revcor data of Carney et al. (1999)
Filterbank structure (a) Signal Input (b) Linear Gammatone Filterbank Output (A) Linear Gammatone Filterbank (c) Linear Gammachirp Filterbank Output (B) Linear Asymmetric Compensation Filterbank (d) Gammachirp Filterbank Output (C) Asymmetric Compensation Filterbank (D) Parameter Controller IIR gammatone IIR Asymmetric Compensation (Slaney, 1993) Filter (Irino & Unoki, 1999)
Current work (1) Cross frequency parameter constraints in the gammachirp filterbank Fitted to human masking data of Rosen, Baker and Darling (1998) Left tails are a little high. 500, 2 k, 4 k It is possible to reduce the rms error. Note: 250, 1 k, 3 k, and 6 k. Hz data omitted for clarity
Current work (2) Constructing a parameter controller • Analytical type (Irino & Unoki, 1999) (B), (C) Asymmetric Compensation Filterbank (D) Parameter Controller Varying Param. c k-th K-th Parameter Control Unit LI Convert LI Activity k-th to LI Σ Para. LI meter LI Weighting (A) Linear Gammatone Filterbank from adjacent channels • Physiological type (A) Linear Gammatone Filterbank (B) Linear Asymmetric Compensation Filterbank (C) Asymmetric Compensation Filterbank (D) Parameter Controller Varying center freq.
Summary Physiological gammachirp filter » Consistent with physiological data – Level-independent chirp » Excellent fit to human masking data – Level-dependent gain and filter shape A physiological gammachirp filterbank » Enable us to simulate an active BM
Mathematical presentation Physiological Gammachirp Passive gammachirp Peak Frequency Asymmetric function Center Frequency Level-dependent Level-independent
IIR Asymmetric Compensation Filter 4 cascaded 2 nd order IIR filter: Symmetrically placed poles and zeros, smaller than 1
Good Approximation ー Original Gammachirp -- IIR Gammachirp rms error: 0. 68 d. B (90 data pairs) ー Asymmetric function -- IIR Asymmetric Compensation filter
ポスターの構成 A 3 X 15(1. 6 m. X 1. 6 m) Title (p. 1) ここをメインに History (p. 2 -5) Physiological Gammachirp (p. 6 -10) Filterbank (p. 11) Future work (p. 12 -14) Mathematical Presentation (p. 15) (or p. 15 -17)
- Slides: 18