Probing Discrete Units in Unit-Based Speech-to-Speech Translation: A Case Study on Central Kurdish

Read full text Summarise with VIRAT
Authors: Lu Zuo, Mohammad Mohammadamini, Aghilas Sini
Cited by: 0
References: 20

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper presents the first systematic study of Speechto-Unit Translation (S2UT) for Central Kurdish (CKB), a low-resource Indo-Iranian language. While S2UT has excelled in high-resource settings, its generalizability to distant languages remains an open question. We evaluate different encoder architectures and target-speech synthesis strategies to bridge the data scarcity gap. Our results reveal a critical dependence on acoustic consistency: while pretrained encoders and single-speaker target data enable fluent translation, multi-speaker configurations trigger "phonetic chaos" during k-means quantization. Through phonetic mapping and latent space visualization, we provide empirical evidence that speaker variability disrupts the formation of a stable phonetic level representation. These findings clarify the mechanistic necessity of speaker normalization in unit-based models and establish a robust benchmark for expanding S2ST to broader low-resource linguistic landscapes.

Fields of Study & Subjects

Speech Recognition and Synthesis Speech and Audio Processing Phonetics and Phonology Research

Publication details

Journal / source
HAL
Publication date
16 April 2026
Language
EN