🧬 biology

AI-Driven Lumped-Element Modeling of Human Respiratory System for Studying Voice Mechanics

This paper presents a novel AI-driven, physics-based lumped-element model that integrates deep learning-extracted vocal fold dynamics with a spring-damper-mass representation of the respiratory system to simulate voice production and predict non-invasively measurable parameters like subglottal pressure and energy transfer mechanisms.

Original authors: Maruf Md Ik, Maryam Naghibolhosseini, Mohsen Zayernouri

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maruf Md Ik, Maryam Naghibolhosseini, Mohsen Zayernouri

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your voice isn't just a sound coming from your throat, but a complex machine where your lungs, windpipe, and vocal cords work together like a team of engineers, a spring-loaded piston, and a pair of flapping doors.

This paper introduces a new "digital twin" of the human voice system. Instead of just looking at the vocal cords in isolation, the researchers built a computer model that simulates the entire voice production chain: from the air leaving the lungs, through the flexible windpipe, to the vibrating vocal cords, and finally out of the mouth.

Here is how they broke it down using simple analogies:

1. The Engine Room: The Lungs as a Piston

Think of your lungs not as a bag of air, but as a spring-loaded piston inside a cylinder (like a syringe).

The Spring: When you breathe out, your lung tissue acts like a rubber band snapping back, pushing the air out.
The Damper: The tissue also has a "sponge-like" quality that slows things down and absorbs energy (viscosity).
The Model: The researchers turned this biological process into a math equation where a heavy weight (the lung mass) is pushed by a spring and slowed by a damper. This allows the computer to calculate exactly how much pressure builds up inside the chest, which is something you can't measure directly inside a living person without surgery.

2. The Flexible Pipes: The Compressible Airways

The windpipe and smaller airways aren't rigid plastic tubes; they are flexible and can squish or expand.

The Analogy: Imagine the airways as collapsible garden hoses. As air rushes through, the walls of the hose can wiggle and change shape.
The Model: The team modeled these as two parallel "pistons" that can move back and forth. This captures how the airways expand and contract, adding a layer of realism that older, simpler models missed.

3. The Doors: The Vocal Cords

The vocal cords are the gatekeepers.

The Analogy: Think of them as a pair of swinging doors that slam shut and fly open hundreds of times a second.
The Data Source: To make the doors move realistically, the researchers didn't just guess. They used AI (Deep Learning) to watch high-speed videos of a real person's throat. The AI acted like a super-fast camera operator, tracing the exact shape of the opening and closing of the vocal cords thousands of times per second.
The Result: The computer model uses these real-life door movements to simulate how air is chopped up into sound waves.

4. The Whole System: A Connected Chain

The magic of this paper is that they connected all these parts into one chain:

Lungs push air (like a piston).
Airways wiggle and resist the flow (like flexible hoses).
Vocal Cords chop the air (like swinging doors).
Mouth releases the air.

Because they are all connected, the model shows how a change in one part affects the others. For example, if the vocal cords slam shut, the pressure builds up in the lungs behind them, which then pushes harder on the next breath.

What Did They Discover?

By running this simulation, the researchers were able to see things that are usually invisible:

The "Invisible" Pressure: They could map the exact pressure changes happening below the vocal cords (subglottal pressure). In real life, you can't stick a pressure gauge there without hurting the person. The model showed that this pressure isn't a smooth, steady line; it rises and falls in a specific pattern that matches the rhythm of the vocal cords.
The "Inverted" Wave: They found that while the air flow (how much air moves) looks like a series of spikes (up and down), the pressure inside the chest looks like the opposite: it stays high and then drops sharply. It's like a dam holding back water; the pressure is high when the gate is closed, and drops when the gate opens.
Realism over Perfection: The model isn't perfect. It assumes the vocal cords close completely and instantly, which is a simplification. In reality, the closing is a bit more gradual and 3D. However, the model successfully captured the main "heartbeat" of the voice system.

Why Does This Matter?

The paper claims this is the first time a model has successfully simulated the respiratory system (lungs and airways) specifically to study how we make voice.

Previously, scientists often just assumed the air coming from the lungs was a steady, boring stream (like a constant breeze). This paper proves that the air is actually a dynamic, pulsing force that changes with every breath and every vibration of the vocal cords. By understanding this "engine" better, we get a more accurate picture of how the human voice actually works, using real data from a healthy person as the baseline.

Technical Summary: AI-Driven Lumped-Element Modeling of Human Respiratory System for Studying Voice Mechanics

Problem Statement
Human voice production relies on the tightly coupled interactions of three subsystems: respiratory (lungs and airways), phonatory (vocal folds), and articulatory (supraglottal vocal tract). Historically, computational models of phonation have focused primarily on the phonatory and articulatory subsystems, often simplifying the respiratory input as a uniform, steady, or sinusoidal subglottal pressure/velocity. These simplifications fail to capture the subject-specific, nonlinear, and irregular dynamics inherent in true respiratory mechanics. Consequently, existing models overlook how specific inlet conditions—shaped by lung tissue viscoelasticity, airway compressibility, and flow resistance—alter jet asymmetry, turbulence, and acoustic loading. There is a critical need for a holistic modeling framework that generates subject-specific, time-dependent boundary conditions for the larynx to accurately simulate voice mechanics.

Methodology
The authors developed a unified, predictive physics-based model that integrates the respiratory, phonatory, and articulatory subsystems into a single lumped-element framework (LEM).

Respiratory Subsystem: The lungs and compressible airways are modeled as spring-damper-mass controlled piston-cylinder systems.
- Lungs: Represented as a single equivalent cylindrical chamber where volumetric changes are governed by piston motion. The model incorporates nonlinear viscoelastic properties of lung tissue via a nonlinear spring (elastic recoil) and both linear and nonlinear damping elements (tissue viscosity and friction).
- Compressible Airways: Modeled as two parallel piston-cylinder assemblies representing peripheral airways and primary bronchi, accounting for wall elasticity and viscous dissipation.
- Flow Resistances: The model quantifies pressure losses across distinct anatomical segments, including lung tissue deformation ( $R_{LT}$ ), small airways ( $FR_s$ ), compressible airways ( $FR_c$ ), trachea ( $FR_T$ ), vocal folds ( $FR_{VF}$ ), upper airways ( $FR_U$ ), and the oral outlet ( $FR_E$ ).
Phonatory Subsystem: The vocal folds are represented using a single-mass model. Crucially, the glottal area waveform (GAW) is not assumed but extracted from laryngeal high-speed videoendoscopy (HSV) data of a normophonic participant using a deep learning-based segmentation framework (U-Net). This data drives the vocal fold dynamics.
Articulatory Subsystem: Represented as the upper airways extending to the oral cavity, accounting for aerodynamic losses prior to sound radiation.
Numerical Implementation: The coupled dynamics are governed by a nonlinear force balance equation and volumetric continuity constraints. The system is solved using a fourth-order Runge-Kutta (RK4) scheme with a time step of $2.5 \times 10^{-4}$ s, matching the resolution of the HSV data. Particle swarm optimization was employed to estimate lung damping coefficients.

Key Contributions

Holistic Integration: This work presents the first LEM that explicitly bridges pulmonary airflow models with phonatory dynamics, extending beyond traditional breathing-only models to simulate sustained phonation.
Data-Driven Boundary Conditions: By integrating AI-driven extraction of GAW from HSV data, the model generates subject-specific, time-dependent inlet conditions rather than relying on idealized sinusoidal or steady inputs.
Nonlinear Viscoelastic Modeling: The framework captures the nonlinear viscoelastic properties of lung tissues and compressible airways, providing a responsive baseline model capable of extension to patient-specific scenarios.
Simulation of Unmeasurable Variables: The model enables the simulation of subglottal pressure distributions and dynamic resistances that are impossible to measure directly and noninvasively in humans during phonation.

Results
The unified model successfully simulated the time-resolved dynamics of a sustained phonation task (sustained /i/):

Glottal Dynamics: The model reproduced quasi-periodic glottal opening and closure, with peak glottal areas of 8–9 mm² and complete closure (zero area) in each cycle. Glottal flow rates ranged from 200 to 240 ml/s, exhibiting stable cycle-to-cycle behavior.
Pressure Dynamics:
- Subglottal Pressure ( $P_{sub}$ ): Oscillated between 400 Pa and 880 Pa. The waveform featured extended pressure plateaus during glottal closure (energy storage) followed by rapid drops during opening.
- Intrapleural Pressure ( $P_{pl}$ ): Ranged from 700 Pa to 1090 Pa, showing a similar but more pronounced pattern to subglottal pressure, reflecting the mechanical load on the lungs.
Flow Resistance: The expiratory flow resistance ( $FR_E$ ) exhibited a pulse-like behavior, rising to peaks of 25–29 $\times 10^3$ Pa/(m³·s) during high flow and collapsing to zero during glottal closure.
Validation: The model was validated against excised-larynx experimental data (red deer larynx). While a localized discrepancy of 13.33% occurred near peak glottal opening (due to the model's inability to capture 3D jet separation), the mean cycle-averaged error was only 2.96%, confirming the model's ability to capture dominant pressure dynamics.

Significance and Claims
The paper claims to establish a foundational benchmark for subject-specific phonation modeling. By transcending the limitations of idealized inlet conditions, this framework advances the prediction of subglottal pressure distributions, dynamic resistances, and energy transfer mechanisms during voice production. The authors position this work as a necessary step toward clinically meaningful computational models that can be adjusted for various pathologies affecting voice production. While the current model relies on 2D glottal area measurements and a single-mass vocal fold representation (limiting its ability to capture 3D motion nuances and fine-scale pressure oscillations during closure), it provides a physiologically meaningful approximation of respiratory mechanics. The authors note that future work will aim to incorporate distributed vocal-tract acoustics, power-law rheology, and more detailed 3D effects to further refine the model.