AES Europe 2026

Current deep learning approaches to speech enhancement rely
heavily on objective measures like mean squared error or
scale-invariant signal-to-distortion ratio as both training
objectives; evaluation metrics. While analytically
convenient, these benchmarks often fail to capture the
nuances of human perception or actual intelligibility.
Furthermore, the inconsistent integration of metrics like
Short-Term Objective Intelligibility or Perceptual
Evaluation of Speech Quality into training; evaluation
pipelines leaves a gap between algorithmic performance;
perceptual reality. This paper proposes a transition
towards evaluation methodologies grounded in
psychoacoustics; audiological modeling. Our study
explores two distinct methods to characterise enhanced
signals. On one hand, we employ a perceptual approach based
on the Cambridge loudness model to assess the preservation
of spectral excitation patterns; perceived intensity. On
the other hand, we adopt a biophysical approach by
utilising CoNNear, a convolutional model of the human
auditory periphery. This allows us to simulate
representations of responses at different stages of the
auditory periphery to observe how speech enhancement
processing affects the physiological representation of
speech. We analyse pre-trained speech enhancement models
using automatic speech recognition; Short-Term Objective
Intelligibility as an additional proxy for human
intelligibility. By mapping automatic speech recognition
performance against loudness; peripheral response
patterns, we investigate the extent to which current
enhancement strategies maintain the perceptual;
physiological integrity of the speech signal. This work
aims to identify features predictive of intelligibility,
providing a foundation for speech enhancement systems
optimised for the human listener rather than purely
signal-based objective functions.