A high accuracy nonlinear model of the human cochlea

cochlear modeling;signal processing;filter fitting;resonant tectorial membrane;audibility;speech perception

Sullivan, Christopher L ; Allen ; Jont B.

关键词: cochlear modeling;    signal processing;    filter fitting;    resonant tectorial membrane;    audibility;    speech perception;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/92668/SULLIVAN-THESIS-2016.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Reliably modeling the human auditory system is of fundamental importance to audio processing systems and hearing research. Generally, models intended for real-time audio processing are time-efficient but tend to lack grounding in physical reality, while models designed for hearing research may closely fit experimental data but cannot always be meaningfully applied in audio processing situations. The goal of this research is to design a computational model of the human auditory system which manages the trade-off between physical correctness and audio processing practicality. The proposed model is a bank of nonlinear digital filters followed by models of the outer and inner hair cells. Methods are introduced which allow for convex optimization of the parameters of the nonlinear filter bank to fit frequency responses generated by a high-accuracy physical model of the auditory system (the Sen-Allen model). Further optimization methods are introduced which fit the parameters of the hair cell models using experimental data on the basilar membrane compression curve and the intensity just-noticeable-difference. The result is an efficient multi-rate system which can be easily reconfigured based on the needs of the application. Preliminary tests of the model show that it is capable of reproducing documented psychoacoustical effects such as pure tone forward and simultaneous masking. Furthermore, an audibility prediction system based on the model is developed and compared to the state-of-the-art articulation index gram. After a brief investigation, the novel system (termed the cochlear voltage difference gram) seems to predict the audibility of speech cues in noise as well as or better than the articulation index gram in most cases, although a thorough comparative analysis must still be conducted. At a 16 [kHz] sampling rate, simulating 100 frequency channels on the cochlea, a Matlab implementation of the model runs about half as fast as real time. Due to the highly parallel nature of the model, it is expected that a similar implementation on a digital signal processor or graphics processing unit could be optimized to run in real time.