End-to-End Neural Position Tracking and Filter Synthesis for Sound Zone Control
Author
Han, Zexu
Term
3. semester
Education
Publication year
2025
Abstract
This report investigates a deep learning-based approach to adaptive Sound Zone Control (SZC), where different listening areas within the same room receive individualized audio without using headphones. Classical SZC techniques such as Pressure Matching (PM) and Acoustic Contrast Control (ACC) rely on static conditions and accurately measured impulse responses, making them sensitive to changes in listener position and acoustic paths and computationally expensive when filters must be updated frequently. Existing solutions either use large dictionaries of precomputed filters selected via position tracking, or compute new filters online in real time, both suffering from limitations in memory, complexity, and dependence on additional tracking hardware. The project instead proposes a neural network that directly maps observation microphone signals to loudspeaker control filters, avoiding explicit position estimation and repeated PM/ACC optimization. During training, physics-inspired loss functions derived from PM and ACC formulations encourage the model to reproduce the desired sound field in the bright zone while suppressing energy in the dark zone. The method is evaluated in scenarios with moving listeners and mismatched impulse responses, and is compared against fixed-filter and audio-based dictionary baselines in terms of robustness, latency, and spatial control. Although detailed numerical results are not included in this excerpt, the work aims to demonstrate that end-to-end neural filter generation from acoustic observations is a promising direction toward real-time, robust SZC in dynamic acoustic environments.
Denne rapport undersøger en dyb-læringsbaseret metode til adaptiv Sound Zone Control (SZC), hvor forskellige lytteområder i samme rum får hver deres lyd uden brug af hovedtelefoner. Klassiske SZC-metoder som Pressure Matching (PM) og Acoustic Contrast Control (ACC er afhængige af statiske omgivelser og præcise forhåndsmålinger af impulssvar, hvilket gør dem sårbare over for ændringer i lytterens position og akustiske forhold samt giver høj beregningskompleksitet ved løbende filteropdatering. Eksisterende løsninger arbejder enten med store ordbøger af præberegnede filtre, der vælges ud fra positionssporing, eller med online beregning af nye filtre, begge med begrænsninger i hukommelsesforbrug, kompleksitet og krav til ekstra sensorer. Projektet foreslår i stedet et neuralt netværk, der direkte kortlægger signaler fra observationsmikrofoner til kontrolfiltre for højttalerarrayet, uden eksplicit positionsestimering eller gentagen PM/ACC-optimering. Under træning anvendes fysik-baserede tabsfunktioner udledt af PM- og ACC-formuleringer, så modellen lærer både at genskabe den ønskede lyd i bright zone og undertrykke energi i dark zone. Metoden undersøges i scenarier med bevægende lyttere og afvigelser mellem præmålt og faktisk impulssvar, og sammenlignes med faste filtre og audio-baserede ordbogsmetoder med hensyn til robusthed, latenstid og rumlig kontrol. Resultaterne (som kun skitseres i uddraget) sigter mod at vise, at end-to-end neuralt filterdesign fra mikrofonobservationer er en lovende vej til mere robust og realtids-egnet SZC i dynamiske akustiske miljøer.
[This abstract has been generated with the help of AI directly from the project full text]
Other projects by the authors