Glimpse Proportion Maximization for Speech Intelligibility Enhancement
Author
Term
4. semester
Education
Publication year
2025
Submitted on
2025-08-25
Pages
45
Abstract
This thesis investigates the use of Glimpse Proportion (GP) maximisation as an optimisation objective for Near-End Listening Enhancement (NELE). Unlike conventional methods based on the Speech Intelligibility Index (SII) or related metrics, the proposed approach employs a differentiable formulation of GP, enabling gradient-based optimisation under energy-preservation constraints. The method, introduced as GlimpseP, applies frequency-dependent, timeinvariant spectral weighting and is evaluated across multiple datasets (DANTALE II, AEMST, TIMIT) and noise conditions (stationary and competing speaker). Results show consistent improvements in objective intelligibility metrics, with particular advantages in fluctuating noise where glimpsing cues are most perceptually relevant. Compared to established baselines such as FractileASII, the proposed method demonstrates comparable or superior performance while maintaining robustness across conditions. These findings confirm the potential of GP as a perceptually grounded optimisation target for NELE.
Keywords
Documents