Andreas G. Andreou
Moise H. Goldstein, Jr.
Dept. of Elec. and Comput. Eng., Johns Hopkins Univ., Baltimore, MD 21218
Uniform bandwidth spectral analysis methods, such as spectrogram, are not suitable for rapidly changing broadband speech signals since the tradeoff between temporal and spectral resolutions is fixed for all frequencies. Described here is the signal representation by a model of the auditory periphery, based on neurophysiological data and realized in analog integrated circuitry. The main modules in the model are the basilar membrane filter bank and the reservoir model of hair cells/synapses [earlier version described in Liu et al., IEEE Trans. Neural Nets 3, 477--487 (1992)]. The instantaneous firing rates of the auditory-nerve fibers are modeled. Although nonconstant Q conditions in the cochlear filter are necessary in order to match the model output to the neural data, the model does exhibit properties of a wavelet analysis. Temporal analysis of the model output yields accurate determination of both time and frequency of each component in synthetic and natural speech signals. Finally, the low-power large-scale integrated circuits operate in real-continuous time (equivalent of 2.8x10[sup 9] arithmetic operations/second), and all model parameters are electronically tunable.