Situated Hearing Aid Personalization as Bayesian Inference

Code repository for the paper:

R. Tresor, M. Hidalgo Araya, W. Nuijten, T. van de Laar, B. de Vries. "Situated Hearing Aid Personalization as Bayesian Inference," 2026.

Overview

This repository implements the validation experiments described in the paper. A Bayesian hearing-aid agent learns user preferences through an online scanning interaction protocol, using variational message passing on Forney-style factor graphs for both preference inference and active trial design.

Two simulation experiments are provided:

1. Inference performance (Section 5.1) -- evaluates the agent under model-matched conditions across a full factorial parameter sweep.
2. Hearing-aid personalization (Section 5.2) -- evaluates the agent in a realistic HA tuning scenario using PESQ as a surrogate for user-perceived quality.

Repository structure

.
├── src/
│   └── common.jl                  # Shared type definitions and utilities
├── experiment/
│   ├── inference_performance.jl   # Experiment 1: inference performance (Sec. 5.1)
│   ├── hearing_aid_personalization.jl  # Experiment 2: HA personalization (Sec. 5.2)
│   └── data_generation.jl         # Synthetic data generation demo
├── environment/
│   ├── environment.jl             # RxEnvironments interaction logic (Exp. 1)
│   ├── user.jl                    # Simulated user model (HAUser)
│   └── agent.jl                   # (Consolidated into src/common.jl)
├── model/
│   └── happiness_tracking.jl      # Generative model: RxInfer @model definitions
├── nodes/
│   ├── extended_probit.jl         # Custom ExtendedProbit factor node
│   └── projection_scheme.jl       # Projection scheme for message passing
├── sound_processing/
│   ├── world.jl                   # RxEnvironments world for audio experiment (Exp. 2)
│   ├── hearingaid.jl              # Warped filterbank hearing aid model
│   └── pnr.jl                     # Power-based noise reduction algorithm
├── metrics/
│   └── metrics.jl                 # PESQ and STOI speech quality metrics (via PyCall)
├── data/
│   ├── clean.wav                  # Clean speech reference (VoiceBank corpus)
│   ├── noisy.wav                  # Noisy speech signal (VoiceBank corpus)
│   └── README.md                  # Data provenance and citation
├── results/                       # Pre-computed experiment results (CSV)
├── showcase.ipynb                 # Analysis and visualization notebook
├── Project.toml                   # Julia dependency specification
└── requirements.txt               # Python dependency specification

Prerequisites

• Julia 1.10+ (tested with 1.10.10)
• Python 3.9+ with packages listed in requirements.txt (only needed for Experiment 2)

Setup

1. Install Julia dependencies

using Pkg
Pkg.activate(".")
Pkg.instantiate()

2. Install Python dependencies (Experiment 2 only)

pip install -r requirements.txt

3. Configure PyCall (if needed)

If Julia's PyCall does not find the correct Python environment:

ENV["PYTHON"] = "/path/to/your/python"
using Pkg; Pkg.build("PyCall")

Reproducing results

Experiment 1: Inference performance (Section 5.1, Table 1)

julia --project=. experiment/inference_performance.jl

Produces: results/experiment_results_2_scannings.csv

Paper mapping: Table 1 -- RRMSD, RMSD_post, and variance ratio metrics across parameter sweep (6,300 runs).

Note: The sleep(1) calls in the experiment loop add significant wall-clock time to batch runs. These can be commented out without affecting results.

Experiment 2: HA personalization (Section 5.2, Table 2 + Figure 7)

julia --project=. experiment/hearing_aid_personalization.jl

Produces: results/experiment_results_audio.csv

Paper mapping: Table 2 -- RRMSD of inferred G_min from oracle optimum; Figure 7 -- scanning trajectories and posterior evolution.

Analysis notebook

The showcase.ipynb notebook loads the result CSVs and generates the tables and figures reported in the paper. To run it:

using Pkg; Pkg.add("IJulia")
using IJulia; notebook()

Script-to-paper mapping

Script	Paper Section	Output
experiment/inference_performance.jl	Sec. 5.1	Table 1 (RRMSD, RMSD, Var. ratio)
experiment/hearing_aid_personalization.jl	Sec. 5.2	Table 2 + Figure 7
experiment/data_generation.jl	Sec. 4 (model demo)	Synthetic data
showcase.ipynb	Sec. 5.1 & 5.2	All tables and figures

Data

Audio files in data/ are excerpts from the VoiceBank corpus:

Valentini-Botinhao, C. (2017). Noisy Speech Database for Training Speech Enhancement Algorithms and TTS Models. University of Edinburgh. DOI: 10.7488/ds/2117

See data/README.md for details.

Key dependencies

• RxInfer.jl -- Bayesian inference via variational message passing
• RxEnvironments.jl -- Reactive environment framework for agent-based simulation

Citation

@article{tresor2025situated,
  title   = {Situated Hearing Aid Personalization as Bayesian Inference},
  author  = {Tresor, Raphael and Hidalgo Araya, Marco and Nuijten, Wouter
             and van de Laar, Thijs and de Vries, Bert},
  year    = {2025}
}