Computationally Efficient DNN-Based Approximation of an Auditory Model for Applications in Speech Processing

Abstract

Computational models of the auditory periphery are important tools for understanding mechanisms of normal and impaired hearing and for developing advanced speech and audio processing algorithms. However, the simulation of accurate neural representations entails a high computational effort. This prevents the use of auditory models in applications with real-time requirements and the design of speech enhancement algorithms based on efficient bio-inspired optimization criteria. Hence, in this work we propose and evaluate DNNbased approximations of a state-of-the-art auditory model. The DNN models yield accurate neurogram predictions for previously unseen speech signals with processing times shorter than signal duration, thus indicating their potential to be applied in real-time.

A comparative study of eight human auditory models of monaural processing

“…The number of signals that can be processed in parallel will depend on the number of threads of the host computer. As a further solution to the long processing time, Stages 2–5 of verhulst2018 (transmission-line, IHC, and AN modules) and bruce2018 (generating mean PSTHs) have been approximated using deep neural networks in [ 99 , 100 ] and [ 101 ], respectively.…”

“…It is important to note that the simplification of auditory models based on statistical methods or machine learning processes requires a careful interpretation. While these approaches might be well suited to achieve goals such as real-time processing (e.g., [ 99 ]) in applications of speech perception (e.g., [ 101 ]) or in the prediction of evoked potentials [ 89 ], they limit the modular comprehension of each auditory stage, especially if multiple model stages are approximated (as in [ 89 , 101 ]).…”

“…When using large datasets, where the stimuli are split into training and validation data (e.g., [ 16 , 101 , 112 ]), the stimuli should contain representative samples of the relevant sound properties that the model user wishes to test. A practice like this can help to support (or not) the applicability of a specific model to sounds that may have not been even validated before, the “unseen sounds”, reducing (or generating awareness of) the potential limitations of the test model.…”

A Non-Intrusive Speech Quality Evaluation Method Based on the Audiogram and Weighted Frequency Information for Hearing Aid

“…DNN can simulate the ability of the human brain to process information and learn the deep characteristics of input signals [21] [22]. Therefore, a DNN module is employed to extract the information related to hearing loss from the audiogram.…”