Applying Biophysical Auditory Periphery Models for Real-time Applications and Studies of Hearing Impairment

Autor: Van Den Broucke, Arthur, Drakopoulos, Fotios, Baby, Deepak, Verhulst, Sarah
Přispěvatelé: Universiteit Gent = Ghent University [Belgium] (UGENT)
Jazyk: angličtina
Rok vydání: 2020
Předmět:
Zdroj: Forum Acusticum
Forum Acusticum, Dec 2020, Lyon, France. pp.3005-3006, ⟨10.48465/fa.2020.1026⟩
PROCEEDINGS OF FORUM ACUSTICUM
e-Forum Acusticum 2020
ISSN: 2221-3767
Popis: International audience; Biophysically realistic models of the cochlea are based on cascaded transmission-line (TL) models which capture longitudinal coupling, cochlear nonlinearities, as well as the human frequency selectivity. However, these models are slow to compute (in the order of seconds/minutes), explaining why less-accurate model descriptions of cochlear processing (e.g., gammatone, DRNL, MFCC) are still the standard for feature extractors or auditory front-ends. To overcome this gap, we present a hybrid approach in which convolutional neural network (CNN) techniques are combined with computational modelling to yield a real-time model of the human auditory periphery. A CNN was trained on speech corpus material to mimic a state-of-the-art biophysical model that can accurately represent the human cochlea and the ascending auditory pathway. The performance was compared against human data and simulations of the original model using basic stimuli (pure tones, clicks, etc.). Because the original peripheral model can simulate different degrees of sensorineural hearing loss, the normal-hearing CNN model can be adjusted in the same fashion to simulate hearing impairment. The neural-network character of these architectures allows for real-time, parallel and differentiable computations, which can serve in the next generation of hearing-aid and machine-hearing applications. Work supported by European Research Council ERC-StG-678120 (RobSpear)
Databáze: OpenAIRE
Externí odkaz: