Showing 1–1 of 1 results for author: Salais, L

BWSNet: Automatic Perceptual Assessment of Audio Signals

Authors: Clément Le Moine Veillon, Victor Rosi, Pablo Arias Sarah, Léane Salais, Nicolas Obin

Abstract: This paper introduces BWSNet, a model that can be trained from raw human judgements obtained through a Best-Worst scaling (BWS) experiment. It maps sound samples into an embedded space that represents the perception of a studied attribute. To this end, we propose a set of cost functions and constraints, interpreting trial-wise ordinal relations as distance comparisons in a metric learning task. We… ▽ More This paper introduces BWSNet, a model that can be trained from raw human judgements obtained through a Best-Worst scaling (BWS) experiment. It maps sound samples into an embedded space that represents the perception of a studied attribute. To this end, we propose a set of cost functions and constraints, interpreting trial-wise ordinal relations as distance comparisons in a metric learning task. We tested our proposal on data from two BWS studies investigating the perception of speech social attitudes and timbral qualities. For both datasets, our results show that the structure of the latent space is faithful to human judgements. △ Less

Submitted 21 January, 2024; v1 submitted 5 September, 2023; originally announced September 2023.