Perceptual similarity between piano notes: Simulations with a template-based perception model

Abstract

In this paper the auditory model developed by Dau et al. [J. Acoust. Soc. Am. 102, 2892-2905] was used to simulate the perceptual similarity between complex sounds. For this purpose, a central processor stage was developed and attached as a back-end module to the auditory model. As complex sounds, a set of recordings of one note played on seven different pianos was used, whose similarity has been recently measured by Osses et al. [J. Acoust. Soc. Am. 146, 1024-1035] using a 3-AFC discrimination task in noise. The auditory model has several processing stages that are to a greater or lesser extent inspired by physiological aspects of the human normal-hearing system. A set of configurable parameters in each stage affects directly the sound (internal) representations that are further processed in the developed central processor. Therefore, a comprehensive review of the model parameters is given, indicating the configuration we chose. This includes an in-depth description of the auditory adaptation stage, the adaptation loops. Simulations of the similarity task were compared with (1) existing experimental data, where they had a moderate to high correlation, and with (2) simulations using an alternative but similar background noise to that of the experiments, which were used to obtain further information about how the participants' responses were weighted as a function of frequency.

A comparative study of eight human auditory models of monaural processing

“…1c ). The linear effective models are represented by dau1997 [ 31 ] and osses2021 [ 39 ] and the nonlinear effective models are represented by king2019 [ 37 ] and relanoiborra2019 [ 38 ]. Given that for each model a similar level of approximation has been generally used in the design of subsequent model stages, we use the defined categories to reflect the nature of the entire model.…”

“…In king2019, adaptation is simulated by applying a highpass filter with a cut-off frequency of 3 Hz [ 37 ]. In dau1997, relanoiborra2019, and osses2021, adaptation is simulated by so-called adaptation loops [ 44 ] that introduce a nearly logarithmic compression to stationary input signals and a linear transformation for fast signal fluctuations (Appendix B in [ 39 ]). The arbitrary units of these transformed outputs are named model units (MUs).…”

“…For these latter models, which include mostly level-independent stages, such calibration is relevant because the adaptation loops (used in dau1997, osses2021, also extensible to relanoiborra2019) include level-dependent scaling (Eqs. (B1)–(B3) in [ 39 ]). In king2019, a calibrated knee point (default of 30 dB) is used in its cochlear compression stage (Stage 3).…”

“…2 ) is used as input to a decision back-end based on a signal-detection-theory (SDT) framework, the template-matching approach. This framework, extended to adopt two templates, has been recently validated to account for the perceptual similarity between two sounds using osses2021 [ 39 ].…”

“…For successful simulations, the decision stage should appropriately weight the information contained in the model representations. An analysis of weighted time-frequency representations (time, audio frequency, and/or modulation frequency) can reveal what portions of the simulated responses are more relevant (e.g., [ 39 , 124 ]).…”