Biomimetic cochlea filters : from modelling, design to analogue VLSI implementation

This thesis presents a novel biomimetic cochlea filter which closely resembles the biological cochlea behaviour. The filter is highly feasible for analogue very-large-scale integration (VLSI) circuits, which leads to a micro-watt-power and millimetre-sized hardware implementation. By virtue of such features, the presented filter contributes to a solid foundation for future biologically-inspired audio signal processors. Unlike existing works, the presented filter is developed by taking direct inspirations from the physiologically measured results of the biological cochlea. Since the biological cochlea has prominently different characteristics of frequency response from low to high frequencies, the biomimetic cochlea filter is built by cascading three sub-filters accordingly: a 2nd-order bandpass filter for the constant gentle low-frequency response, a 2nd-order tunable low-pass filter for the variable and selective centre frequency response and a 5th-order elliptic filter for the ultra-steep roll-off at stop-band. As a proof of concept, a biomimetic cochlea filter bank is built to process audio signals, which demonstrates the highly discriminative spectral decomposition and high-resolution time-frequency analysis capabilities similar to the biological cochlea. The filter has simple representation in the Laplace domain which leads to a convenient analogue circuit realisation. A floating-active-inductor circuit cell is developed to build the corresponding RLC ladder for each of the three sub-filters. The circuits are designed based on complementary metal-oxide-semiconductor (CMOS) transistors for VLSI implementation. Non-ideal factors of CMOS transistors including parasitics, noise and mismatches are extensively analysed and consciously considered in the circuit design. An analogue VLSI chip is successfully fabricated using 0.35μ m CMOS process. The chip measurements demonstrate that the centre frequency response of the filter has about 20 dB wide gain tuning range and a high quality factor reaching maximally over 19. The filter has a 20 dB/decade constant gentle low-frequency tail and an over 300 dB/decade sharp stop-band roll-off slope. The measured results agree with the filter model expectations and are comparable with the biological cochlea characteristics. Each filter channel consumes as low as 59.5 ~90μ Wpower and occupies only 0.9 mm2 area. Besides, the biomimetic cochlea filter chip is characterised from a wide range of angles and the experimental results cover not only the auditory filter specifications but also the integrated circuit design considerations. Furthermore, following the progressive development of the acoustic resonator based on microelectro- mechanical systems (MEMS) technology, a MEMS-CMOS implementation of the proposed filter becomes possible in the future. A key challenge for such implementation is the low sensing capacitance of the MEMS resonator which suffers significantly from sensitivity degradation due to the parasitic capacitance. A novel MEMS capacitive interface circuit chip is additionally developed to solve this issue. As shown in the chip results, the interface circuit is able to cancel the parasitic capacitance and increase the sensitivity of capacitive sensors by 35 dB without consuming any extra power. Besides, the chopper-stabilisation technique is employed which effectively reduces the circuit flicker noise and offsets. Due to these features, the interface circuit chip is capable of converting a 7.5 fF capacitance change of a 1-Volt-biased 0.5 pF capacitive sensor pair into a 0.745 V signal-conditioned output while consuming only 165.2μ W power