The information-differentiated loss function for speech feature clustering in a low-resource environment
Year: 2025
Type: conference-paper
Abstract: The article investigates the problem of forming a portable cluster structure in the latent space of language representations without using annotated data, which is especially relevant for zero-shot classification tasks, low-resource language processing, and generalization to new domains.Based on a critical review of modern approaches to unsupervised cluster learning, a loss function is proposed that combines global entropy regularization with scaling of the contribution of examples depending on the level of model confidence.The value of the scaling parameter is determined automatically based on the local decrease in the entropy of the cluster distribution, which serves as an indicator of the isolation of a language segment in the latent space.Such a mechanism makes it possible to suppress the contribution of latently unstable examples without removing them, ensuring structural adaptation of the cluster topology in new domains without retraining.Experimental results on GlobalPhone, CommonVoice and unseen-domain Ukrainian Speech Corpus demonstrated a reduction in average cluster entropy to 0.88, suppression of over 60% of unstable segments and an increase in cluster structure consistency by 19% in zero-shot mode.The proposed approach provides stable and adaptive clustering in the absence of annotations, in particular in cold start scenarios and rapid structuring of speech data in new environments.
Authors: Viacheslav Kovtun, Victoria Vysotska, Oksana Kovtun
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7084098249
FWCI:
Cites:
Cited by:
Related to: