AES Europe 2026

This paper introduces a novel approach for generating a
lower layer in multichannel audio upmixing, addressing a
gap in existing methods that primarily focus on mid; top
layers. Leveraging Harmonic-Percussive Separation (HPS),
the proposed framework dynamically adjusts key parameters
(separation factor, harmonic attenuation,; phase shift)
to enhance percussive components while diffusing harmonic
elements. We compared three neural network architectures
for this task: LSTM, TCN,; Transformer. Experimental
results show comparable perceptual quality; objective
metrics across all models, with the TCN being the most
balanced; suitable for deployment on edge devices.

The mixing stage in music production involves a complex set
of interdependent technical; creative decisions aimed at
achieving a coherent; industry-level result. Intelligent
Music Production (IMP) is an emerging research area that
integrates Artificial Intelligence techniques into music
creation; post-production processes, spanning from
composition to mastering. Within this context, Answer Set
Programming (ASP), a declarative paradigm from Knowledge
Representation; Reasoning, has proven effective for
modeling; solving complex optimization problems. This
article presents frmixerr, an ASP-based intelligent system
designed to optimize the mixing process by automatically
generating balanced mixes. The system formulates mixing as
a combinatorial optimization problem; evaluates
candidate solutions against a reference spectral profile.
To assess its performance, a subjective listening test was
conducted comparing mixes generated by frmixerr with mixes
produced by human engineers with varying levels of
professional experience. The results indicate no
significant differences in perceived quality between
frmixerr mix; those created by professionals, suggesting
that ASP constitutes a viable approach for intelligent
assistance in music mixing.

The development of personal sound zone systems in recent
years show great potential for low-frequency noise control
outside of noisy spaces. These approaches show promising
applications to manage noise pollution arising from
concerts in large venues or urban festivals. However, most
of the literature considered that the created sound zones
would exist in the same room or acoustic space as the noise
source. This premise hence discards all setups where the
disturbances would occur outside of concert venues (e.g in
neighboring houses). This paper presents a first
experimental study of the behavior of sound zone methods
for indoor sound zones; outdoor noise sources. These
initial results present a good efficiency of these methods
in this edge case, opening new use cases for these
approaches.

Conventional ornithological monitoring systems rely heavily
on single-channel recorders; deep learning classifiers
to identify "what" species is present, but fail to capture
"where" it is located or how individuals interact
spatially. This limitation hinders the study of complex
ecological behaviors, such as inter-specific spacing in
dense vegetation; predator-prey dynamics. We propose a
novel, dual-mode acoustic localization system designed to
unify semantic classification; spatial tracking.
Utilizing an economically scalable 16-channel Uniform
Rectangular Array (UMA-16) interfaced with edge-computing
platforms, we implement a hybrid spatial filtering pipeline
structured to balance real-time latency constraints with
achievable angular resolution. The first stage employs a
computationally efficient, noise-robust linear scanning
technique to generate an acoustic energy map; estimate
source multiplicity. This preliminary data initializes a
second-stage, super-resolution spectral estimation
algorithm predicated on signal-noise subspace
orthogonality, allowing the noise robustness of
non-parametric beamforming methods with the precision of
parametric approaches. By integrating these spatial filters
with standard deep learning classifiers, the system
resolves overlapping vocalizations in "Cocktail Party"
scenarios; improves Signal-to-Noise Ratio (SNR) for
cryptic species detection. We address the physical
"Localization-Detection Range Disparity," demonstrating
that while detection is viable at long ranges, precise
localization is constrained by the array aperture to the
near-to-mid field. The system outputs real-time video
overlays of acoustic heatmaps for field observation;
generates autonomous volumetric territory maps in fixed
deployments, collectively providing ornithologists with a
robust capability for analyzing the spatial ecology of
avian vocalizations.

There are many types of different distortions that can be
measured from linear to non-linear distortion. Often the
two are convoluted together and the linear distortion
influences the non-linear distortion. Distortion is also
very signal and level dependent and it is hard to compare
one type of distortion measurement to another. There are
many type of non-linear distortion metrics, e.g. THD, THD+N
and IMD being the most classic ones using sine tones as the
test signal. But how can we measure distortion with real
signals such as speech and music or even noise and compare
the results to audibility? This tutorial discusses a wide
range of distortion measurements, discusses what is audible
and what distortion sounds like.

Before digital signal processing took over electronic
keyboard instruments, they were implemented using analogue
circuits that used tubes/valves, transistors, and even neon
lightbulbs! Yet using these components keyboards were
developed that could mimic string and brass ensembles,
pianos and harpsichords and many other instruments. How did
they do it?

The purpose of this tutorial is to look at both the
architecture and the circuitry of these instruments. And
show how amazing results could be achieved using
comparatively simple electronic circuitry. It will look at:

1. The basic architecture of these instruments
2. How they generated the right notes,
3. How they desired envelope,
4. And imposed them on the waveform,
5. Simulated the effect of many instruments playing
together.

It will also look at how, if it was required, touch
sensitivity could be achieved, such as in electronic
pianos. Where possible there will be audio examples
demonstrating the sounds that could be achieved.

For many people who have only ever experienced the digital
world it will be illuminating to see just how much could be
achieved by comparatively simple circuits.
In those days electronic components were expensive so
considerable ingenuity was expended in minimising the total
number of components required.

These instruments are part of our musical and audio
heritage and the circuit techniques they used are in danger
of being forgotten so this tutorial will be a timely
reminder of what used to be done.
It may also provide useful information to people who are
attempting to model these instruments using modern digital
methods.

The tutorial will be accessible to everyone, you will not
have to be an electronic engineer to understand the
principles behind these unique pieces of audio engineering
history.

Input-output linearization is a technique for compensating
nonlinear distortion in loudspeakers. To apply it to
complex loudspeaker models, we describe an end-to-end
framework for estimating model parameters from data;
deriving the linearizing control laws using automatic
differentiation. The parameter estimation approach combines
frequency-domain linear parameter estimation with a
time-domain prediction-error method for the nonlinear
parameters. The linearization approach supports non-linear
reference systems; stabilization of the control law
using trajectory tracking. We implement the framework in
dynax, an open-source Python package based on JAX,;
validate it experimentally as a feed-forward controller on
a closed-box loudspeaker. Results demonstrate validation
errors of 1--5\,\% NRMSE; total harmonic distortion
reductions of 6--12\,dB. The framework enables researchers
; engineers to rapidly prototype; validate complex
loudspeaker models for distortion compensation without
manual symbolic derivations.

A low-parameter-count machine-learning model for
classifying streaming video can enable content-aware
audio/video processing on consumer edge devices with
latency, computational,; battery constraints. In this
paper, we propose a low-compute classification technique
that uses only text metadata from the streaming file
header, enabling near-instantaneous inference without
decoding; analyzing audio or video signals as is
traditionally done. In particular, to support multilingual
platforms such as YouTube, we first apply neural machine
translation as a pre-processing step for the text metadata
; optimize a lightweight neural classifier for a
three-class audio-centric classification taxonomy (movie,
music, dialog/other). Experiments on a mixed-language
YouTube dataset achieve $\approx$90\% classification
accuracy on a test set using a combined translation; a
classification model (with only $\sim22K$ parameters),
demonstrating a globally-scalable approach for robust
classification on the edge.

Streaming of immersive audio is known to western audiences
almost exclusively in the object-based format, Atmos,
developed by Dolby and employing lossy codecs to limit bit
rates. Other object-based formats like Sony 360 have had
limited success, and until recently there were no channel
based streamed versions. But this situation is changing,
as it has already done in Japan.

Responding to growing interest in very high quality
immersive music for both on-demand streaming and live
broadcast, two new services are now active that offer,
first, channel-based audio and second, audio streamed in
high res PCM. Binaural mixes, a range of PCM formats and
video are variously included, with extensions to portables,
loudspeakers, and home theater.

This workshop provides a forum for discussion of both the
genuine promise and the challenges in these new
initiatives. Included are the advantages of high
resolution over lossy; channel-based versus object-based;
the degree of adoption of transducers for multichannel;
adaptive bit rates; data sources; and the Japanese
approach; amongst others.

We present Binaspect, an open-source Python library for
binaural audio analysis, visualization,; feature
generation. Binaspect generates interpretable “azimuth
maps” by calculating modified interaural time; level
difference spectrograms,; clustering those
time-frequency (TF) bins into stable time-azimuth histogram
representations. This allows multiple active sources to
appear as distinct azimuthal clusters, while degradations
manifest as broadened, diffused, or shifted distributions.
Crucially, Binaspect operates blindly on audio, requiring
no prior knowledge of head models. These visualizations
enable researchers; engineers to observe how binaural
cues are degraded by codec; renderer design choices,
among other downstream processes. We demonstrate the tool
on bitrate ladders, ambisonic rendering,; VBAP source
positioning, where degradations are clearly revealed. In
addition to their diagnostic value, the proposed
representations can be exported as structured features
suitable for training machine learning models in quality
prediction, spatial audio classification,; other
binaural tasks. Binaspect is released under an open-source
license with full reproducibility scripts at: (link removed
for blind review)

Realistic spatial audio consistent with visual information
is essential for providing high immersion in Augmented
Reality (AR) environments. However, conventional
high-precision real-time acoustic simulations require
significant computational power, limiting their
implementation on standalone mobile VR devices such as the
Meta Quest. This study proposes a practical method to
enhance reverb realism using solely a standalone VR HMD,
without the need for additional external equipment. By
measuring impulse responses using a few hand claps in the
physical space, we interpolate room acoustic
parameters—specifically RT60; early/late energy
ratios—to reflect the environment's unique characteristics.
These extracted parameters are then applied to the VR
engine's built-in reverb effects, enabling dynamic,
location-aware real-time rendering with minimal
computational load. The proposed method demonstrates that a
brief calibration period of 3 to 5 minutes yields
significantly improved realism compared to static reverb
templates, offering an efficient; practical spatial
audio solution for mobile
AR environments.

Deep learning has significantly improved speech enhancement
performance in controlled laboratory conditions, yet these
advances rarely translate into robust real-world benefit
for hearing aid users. Current algorithms are trained;
evaluated in simplified acoustic scenarios, neglecting
multimodal cues, user interaction, environmental dynamics,
; the strict latency; power constraints of embedded
devices. As a result, a persistent gap remains between
algorithmic performance; everyday listening experience.
This position paper reviews recent progress in speech
enhancement, embedded Artificial Intelligence hardware,;
hearing aid systems,; argues for a shift toward
ecologically valid evaluation; hardware-aware design. We
propose virtual reality as a reproducible, multisensory
benchmarking platform enabling joint assessment of human
perception; algorithmic processing. This perspective
outlines a research roadmap toward adaptive, context-aware,
; practically deployable hearing technologies.

This work presents a perceptual model based on a complex
IIR filterbank. The filterbank with a frequency resolution
of 4 bands per Bark consists of 104 filters whose slopes
are designed to take spectral masking effects into account.
The filter outputs are used to obtain masking thresholds
with the following post processing. To obtain resonable
masking thresholds from the spreading outputs, a post
masking stage is required. Here, we propose a comodulation
dependent adaptation of the postmasking decay to model
Comodulation Masking Release (CMR) effects. This approach
explicitely considers the dip-listening effect known from
literature. The final masking thresholds are obtained by
weighting the postmasking outputs by a tonality dependent
gain, controlled using spectral flatness estimation. A
listening test compares the proposed method to an already
known approach using direct CMR based modification of the
masking threshold gains.

Sound source localization; identity tracking are
fundamental tasks in acoustic scene analysis, enabling
machines to determine what, where; when produces sound
events. While deep attractor-based networks have
demonstrated improved performance under an unknown number
of sources, maintaining continuous source tracking over
long-form audio remains challenging due to memory
limitations; permutation ambiguities across adjacent
segments. In this paper, we propose a Recursive Attractor
Network (RANet) for long-form sound source localization;
identity tracking with a variable number of sources. RANet
explicitly represents source attractors as transferable
embeddings; recursively propagates them across adjacent
audio segments using a LSTM-based model, thereby preserving
source identity continuity over time. Experimental results
on simulated datasets demonstrate that RANet achieves
robust long-form sound source localization; consistent
source identity tracking, outperforming baseline approaches
under variable; dynamic source conditions.

Identifying robust headphone target curves is challenging
when preference data from untrained listeners are
interpreted without explicit perceptual structure. This
work presents a methodological framework in which deep-
learning-driven sensory-profile analysis serves as the
primary interpretive layer for listening data.
Candidate target curves are generated using an Interactive
Differential Evolution (IDE) listening experiment that
combines paired comparisons with a second- stage
absolute-rating task, enabling continuous exploration of the
perceptually relevant tuning space while reducing cognitive
load. Converged gain sets are analyzed using a Virtual
Listener Panel (VLP), a Deep Learning (DL) model trained on
large-scale expert evaluations to predict perceptual
attributes from rendered musical material. Predicted
attributes are reported as relative scores along key sensory
dimensions, including bass strength, timbral balance,;
brilliance, enabling exploration of sensory clusters,
perceptual trade-offs,; potential families of target
tunings.
Adaptive listening data from three culturally distinct
listener panels (Denmark, Japan,; Colombia; 20
participants
per site) support the DL-based interpretation. Convergence
is quantified as a reduction in population variance,
; cross-site analyses assess the similarity of clustering
structures; the consistency of relationships between
preference; sensory attributes. Overall, the framework
provides a scalable, perceptually grounded approach to
interpreting listener-preference data when developing
headphone target curves.

Sa quintina is a distinctive emergent vocal phenomenon
almost exclusively associated with the sacred polyphonic
singing tradition of Castelsardo, perceived as an
autonomous “fifth voice” arising during collective
performance by four male singers. Although widely
acknowledged in ethnomusicological literature, its
formation mechanisms remain only partially explored within
audio engineering; acoustical research.
This paper presents an early-stage, descriptive sonological
case study proposing new hypotheses on the formation;
spatial reinforcement of sa quintina. The phenomenon is
interpreted as a physically grounded, measurable outcome of
harmonic fusion; spatial interference, observable
through spectral energy distribution; coherence. It is
hypothesized to emerge from a converging set of
conditions—including non-tempered harmonic textures,
differentiated vocal emission techniques, intentional
formant tuning,; circular spatial configuration—none of
which is assumed to be strictly sufficient in isolation.
Building upon previous spectral coherence analyses, the
study introduces a Quintina Directionality Index (QDI) to
quantify the spatial dimension of the phenomenon. QDI is
defined as the ratio between spectral energy in two
frequency bands associated with sa quintina (600–750 Hz;
1200–1400 Hz); total spectral energy. The index is
evaluated as a function of direction using ambisonic
recordings in an anechoic chamber; as a function of
microphone position in a controlled field setting.
Preliminary observations suggest that sa quintina
corresponds to localized regions of enhanced spectral
coherence; energy reinforcement, supporting its
interpretation as an emergent physical phenomenon that
precedes; enables its perceptual salience, rather than a
purely auditory illusion.

This paper presents a method for extracting a center signal
from two-channel stereo signals for upmixing;
reproduction with additional center loudspeakers.
It uses a generative adversarial network with a generator
trained with multiple reconstruction losses; adversarial
losses obtained from a discriminator.
The processing is of low computationally complexity, causal
; can be configured for latencies down to one audio frame
of 46 ms length.
It is described how training data are created using only
publicly available signals; how the generation of target
data enables to control the attenuation of diffuse signals
; direct signals panned off-center.
An evaluation with listening test; computational metrics
SI-SDR; F2 measure is presented.
It shows an advantage compared to methods based on
classical signal processing in terms of computational
metrics for source separation; listeners preference.

Audio engineering often implicitly assumes a uniformity in
hearing across listeners; this is an assumption that does
not reflect real-world diversity. How could technologies
and practices in production, mixing, and reproduction be
adapted to create music that is more inclusive? While the
AES has a conference series on Audio and Music Induced
Hearing Disorders, this has focused on the causes of
hearing loss with little on audio engineering for listeners
who have a hearing loss.

In western countries, about one in three adults are deaf,
have hearing loss or suffer from tinnitus. Hearing loss can
lead to many challenges with music such as: inaudibility of
quieter passages, distortion, degraded pitch perception,
and difficulty in identifying and picking out lyrics and
instruments. The most common intervention for mild to
moderately severe hearing loss is hearing aids. But while
many of these devices have music programs, their efficacy
is mixed, to the point that many opt not to use them. With
the rise of machine learning within Audio Engineering,
there are opportunities to better personalise music, and
therefore address issues listeners face. Consumer devices
are also increasingly having audio accessibility features
added, but the usefulness of these lack independent
testing. This workshop will consider opportunities for
making music more accessible.

The workshop will start by exploring how hearing loss harms
the experience of listening to music and how this varies
between people. This will lead to discussion of why no
technology can fully ‘correct’ music to achieve a ‘perfect’
listening experience for those with hearing loss. There is
no technology to recreate a ‘golden-ears’ experience. This
leads to a key research question: what is the best,
rendition of a piece of music for someone who has hearing
loss? What do listeners want from music, and how can we get
closest to achieving that?

We will bring in findings from research projects and
listening tests to explore what is known, and also to
highlight that there are significant gaps in knowledge that
require further research. We will then explore
state-of-the-art in wearables such as hearing aids and
sound reproduction systems. This will include the current
Cadenza project, which has been running a series of machine
learning challenges to improve music for those with hearing
loss.

Throughout, we will encourage questions and engagement from
delegates. We want to hear about lived experience of
hearing difference and how that has changed professional
practice and personal lives. We are also keen to hear
suggestions from delegates on what approaches might be used
to improve music for those with hearing loss.

We aim to raise awareness of the importance of considering
diverse audiences in Audio Engineering practice. Where
possible, the workshop will provide practical guidance for
audio engineers, highlighting techniques and emerging
technologies that can better support listeners with diverse
hearing profiles.

The Workshop will be organised by the Cadenza Project Team
https://cadenzachallenge.org/ A large UK-funded project
about improving music for those with hearing loss.

This paper presents Part 2 of our study on personalized
timbre optimization for stereophonic sound reproduction via
earphones, following our previous work presented at the AES
International Conference on Headphone Technology in 2025.
While Part 1 established a novel auditory-model-based
framework for reproducing a listener’s natural timbre
reference; demonstrated its perceptual validity under
controlled conditions, the present study focuses on the
practical implementation; validation of this approach
for real-world use with consumer True Wireless Stereo (TWS)
earphones.

Conventional headphone; earphone personalization
techniques primarily target spatial audio reproduction or
rely on preference-based equalization, often overlooking
the accurate reproduction of natural timbre in stereophonic
content. Our approach explicitly addresses this limitation
by isolating; optimizing perceptually relevant timbral
cues while excluding spatial encoding components, thereby
improving timbral fidelity without degrading stereo imaging.

The proposed method originally consists of four stages:
high-resolution anatomical scanning of the listener’s upper
body, including the pinnae, individualized HRTF computation
using the boundary element method, selective removal of
spatial encoding components to derive a personalized
reference target response curve (PR-TRC),; perceptual
optimization using a listener-specific weighting
coefficient grounded in auditory reference fidelity rather
than preference. In this paper, each stage is simplified
; automated using smartphone-based scanning;
AI-assisted processing, enabling end users to complete the
entire personalization process via a smartphone connected
to a cloud-based server. The resulting personalized target
response curve is implemented within the computational;
memory constraints of the DSP pipeline of commercial
consumer TWS earphones.

A subjective evaluation using the Semantic Differential
Method was conducted to assess the perceptual impact of the
simplified implementation. Twenty-four listeners evaluated
personalized target curves generated by both the original
; simplified methods, as well as two non-personalized
target curves commonly used in commercial TWS earphones.
The results show that both personalized methods
consistently outperform non-personalized conditions in
overall sound quality; listener preference. Importantly,
no statistically significant degradation in perceived
timbral naturalness was observed between the simplified;
original methods.

These findings demonstrate that auditory-model-based
personalized timbre optimization can be effectively
translated into a practical, consumer-ready technology. The
proposed approach represents a foundational contribution to
future audio personalization; has broad applicability
across headphone; earphone systems for stereophonic
sound reproduction.

While Neural Audio Codecs (NAC) have revolutionized
monaural audio compression, achieving high-fidelity
dual-channel coding at low bitrates remains a significant
challenge. Existing approaches often rely on naive
independent channel quantization, leading to phase
incoherence, or entangled latent modeling, which sacrifices
spatial precision for spectral energy. This paper proposes
a novel dual-channel coding framework based on
contentspatial disentanglement. Reframing spatial
reconstruction as an informed source separation task, our
architecture synergizes a frozen, pre-trained DAC encoder
for robust mono content preservation with a
parameter-efficient side information encoder that predicts
fine-grained time-frequency masks. To ensure precise
spatial imaging, we introduce explicit physical constraints
into the end-to-end training. Experimental results indicate
that at low bitrates of 9; 11 kbps, the proposed method
outperforms state-of-the-art dual-mono neural baselines;
industry standards in both objective spatial metrics;
subjective MUSHRA evaluations.