⚡ electrical engineering

Short-wave admittance correction for a time-domain cochlear transmission line model

This paper introduces a numerical short-wave admittance correction using autoregressive filtering and regression to integrate higher-dimensional effects into a time-domain cochlear transmission line model, thereby successfully decoupling frequency selectivity from gain and improving the simulation of gerbil cochlear compression.

Original authors: François Deloche, Morgan Thienpont, Sarah Verhulst

Published 2026-02-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: François Deloche, Morgan Thienpont, Sarah Verhulst

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your inner ear (the cochlea) as a long, winding slide where sound waves travel. When a sound enters, it creates a ripple that moves down this slide, getting bigger and bigger until it hits a specific spot where it "pops" (this is where you hear the pitch). Scientists use computer models to simulate this slide to understand how we hear.

For a long time, these computer models worked like a one-dimensional (1D) tube. They were great at showing how the wave moves forward, but they missed a crucial detail: the slide isn't just a flat tube; it's a 3D space where the water (fluid) behaves in complex ways, especially when the ripples get very short and fast.

Here is what this paper does, explained simply:

1. The Problem: The "Flat" Model Was Too Weak

The researchers were trying to build a model for a gerbil's ear (a small rodent). They used a standard 1D model (called the V-1D model).

The Issue: In real gerbil ears, the sound wave gets very loud (compressed) over a wide range of volumes, but the "sharpness" of the hearing doesn't change much.
The Model's Flaw: The 1D model was like a rubber band: if you pulled it to make it louder (gain), it got wider and less sharp. If you made it sharp, it couldn't get very loud. It couldn't handle the gerbil's specific hearing style, resulting in a model that was about 10 decibels too quiet and didn't compress sound enough.

2. The Solution: Adding a "Magnifying Glass"

The researchers knew that in the real 3D world, as the sound wave gets shorter and faster, the fluid pressure gets squeezed into a tiny layer right against the "slide" (the basilar membrane). This is called pressure focusing. It acts like a natural magnifying glass, boosting the signal without making the hearing less sharp.

However, simulating this 3D "magnifying glass" in a fast, 1D computer model is very hard. It's like trying to describe a 3D shadow using only a 2D drawing.

3. The Fix: The "Smart Filter" (The $V^*$ Model)

The team created a new version of the model (called $V^*$ ) that adds a "correction factor." Think of this as a smart digital filter that sits on top of the original model.

How it works: Instead of trying to simulate the complex 3D physics in real-time (which is too slow), they used a clever trick. They ran a slow, perfect 3D simulation first to see exactly how much the "magnifying glass" should boost the sound.
The Shortcut: They then used math regression (a type of pattern-finding) to teach the fast 1D model how to mimic that boost. They created a "lookup table" of instructions. When the sound gets loud, the model checks the table and applies a specific "boost" to the signal, just like the real ear does.

4. The Results: A Better Gerbil Ear

By adding this "smart filter," the new model achieved two things:

More Volume: It added about 5 decibels of extra gain (loudness) at the peak of the wave.
Wider Range: It extended the range where the ear compresses sound by 10 decibels.
The Trade-off: It didn't solve everything. The model still wasn't quite as sharp as a real gerbil ear at high volumes, but it was much closer than before.

5. Why This Matters (The "Causality" Catch)

The paper ends with a fascinating philosophical point about time.

In the real 3D world, the "magnifying glass" effect happens almost instantly as the wave travels.
In the computer model, everything must happen in a strict order (cause then effect).
The researchers found that trying to force this "instant" 3D boost into a strict "cause-and-effect" 1D timeline is mathematically tricky. It's like trying to make a shadow appear before the object casts it. They had to be very careful with their math to ensure the model didn't break or become unstable, while still capturing the feeling of that 3D boost.

Summary

The authors took a simplified, flat model of a gerbil's ear that was too weak and "stiff." They added a mathematical "boost" button (derived from complex 3D physics) that acts like a dynamic magnifying glass. This allowed the computer model to simulate the ear's ability to get louder without losing its sharpness, making it a much better tool for studying how small mammals hear.

Technical Summary: Short-wave admittance correction for a time-domain cochlear transmission line model

Problem Statement
Time-domain transmission line (TL) models are efficient tools for simulating basilar membrane (BM) displacement in response to transient sounds. However, their one-dimensional (1-D) nature limits the integration of higher-dimensional hydrodynamic effects, specifically pressure focusing and transverse viscous damping, which are prominent in the short-wave region of the cochlea. In the standard V-1D model (Verhulst et al., 2018), gain and frequency selectivity are tightly coupled: achieving high gain requires sharp tuning, while broader tuning results in lower gain. This coupling poses a significant challenge when adapting the model to small mammals like gerbils. Experimental data indicates that gerbil cochlear filters exhibit tall peaks with relatively broad tuning at low sound levels, yet maintain a wide range of compressive nonlinearity. The standard V-1D model fails to reproduce this, presenting insufficient compression (approximately 10 dB short) because its 1-D formulation cannot decouple gain from frequency selectivity without compromising physiological realism.

Methodology
The authors propose a modified model, the $V^*$ model, which integrates a numerical correction for short-wave admittance into the existing time-domain V-1D framework. The methodology involves three primary phases:

Reference Modeling (S-2D): The authors utilize the S-2D model (Sisto et al., 2021), a frequency-domain TL model incorporating 2-D fluid effects (pressure focusing and viscous damping). This model provides a "target" for the gain enhancement factor ( $\alpha$ ) resulting from pressure focusing. A normalized correction factor, $\beta = \alpha / \alpha_0$ , is defined relative to a reference active process strength ( $G_{REF}$ ).
Regression and Filter Design: Since $\beta$ $β$ varies with frequency, BM characteristic frequency ( $\omega_{BM}$ $ω_{B M}$ ), and active process strength ( $G$ $G$ ), and cannot be directly implemented as a simple causal time-domain term, the authors approximate $\beta$ $β$ using autoregressive (all-pole) filters.
- Step 1: For fixed pairs of $(\omega_{BM}, G)$ , 32 autoregressive coefficients are determined via penalized least-squares regression to match the frequency response of the target $\beta$ . Penalty terms ensure coefficient stability and prevent high-frequency instability.
- Step 2: A Radial Basis Function Neural Network (RBF-NN) is trained to map the continuous inputs $(\omega_{BM}, G)$ to the set of 32 filter coefficients, enabling dynamic updates during simulation.
Time-Domain Integration: The $V^*$ model incorporates these recursive filters into the BM admittance equations. The filter coefficients are updated dynamically based on a short-term average of the output intensity, which modulates the strength of the active process ( $G$ ). To balance computational cost and accuracy, the authors implemented a lookup table for pre-computed coefficients and selected an update rate of 0.03 ms (6 times slower than the simulation timestep), which was found to be a sufficient trade-off.

Key Results

Decoupling of Gain and Selectivity: The $V^*$ model successfully introduced a degree of decoupling between frequency selectivity and gain. While the V-1D model could not simultaneously achieve the broad tuning and high compression observed in gerbils, the $V^*$ model extended the range of the compressive regime by 10 dB.
Gain Enhancement: The model achieved an additional 5 dB of gain at the wave peak compared to the uncorrected V-1D model. This improvement allowed the model to better match the experimental growth functions of BM velocity in gerbils (He et al., 2022) without requiring unrealistically sharp tuning at low sound levels.
Stability and Accuracy: The autoregressive filters remained stable, with no observed instability in the z-domain. The regression method captured the general trend of the pressure focusing factor, though it slightly underestimated gain enhancements and overestimated gain reductions, indicating the method is best suited for moderate corrections.
Computational Efficiency: While the dynamic update of filter coefficients initially increased computation time by a factor of 20, the use of a pre-computed lookup table reduced simulation times to be comparable to the original V-1D model.

Significance and Claims
The paper claims that the $V^*$ model offers a practical solution for implementing wavelength-dependent effects in time-domain TL models, specifically addressing the limitation of insufficient compression in models for small mammals. The work highlights two main contributions:

Methodological Integration: It demonstrates a hybrid approach combining analytical fluid dynamics (S-2D) with regression-based filter design (RBF-NN and autoregressive filtering) to translate frequency-domain 2-D effects into a time-domain 1-D framework.
Nonlinearity Combination: The model uniquely combines instantaneous nonlinearities (dependent on instantaneous BM velocity) with non-instantaneous nonlinearities (dependent on the signal envelope via the feedback loop for the gain correction). This allows the model to simulate complex hybrid dynamics where pressure focusing acts as a passive phenomenon modulated by the active process.

The authors modestly note that while the $V^*$ model improves the match to experimental data for the basal turn of the gerbil cochlea, it does not fully resolve the "tip-to-tail" ratio discrepancy (the absolute gain remains lower than experimental observations). They attribute this to the fundamental difficulty of implementing high-gain, causal filters in time-domain models, where local causality constraints limit the ability to replicate the global wave causality and instantaneous pressure focusing found in 2-D physical models. The paper concludes that this approach is particularly valuable for translational research involving small laboratory animals, where extensive physiological data is available but difficult to model with standard 1-D assumptions.