In low-resource multilingual speech-to-text translation, uniform architectural sharing across languages frequently introduces representation conflicts that impede convergence. This work proposes a principled methodology to automatically determine layer-specific sharing patterns by mining training gradient information. Our approach employs three distinct analysis strategies: distance-based language clustering, self/cross-task divergence metrics for capacity allocation, and joint factorization coupled with canonical correlation analysis for subspace alignment. Extensive evaluation across four language pairs (using the SeamlessM4T-Medium architecture) demonstrates persistent improvements in translation quality metrics.
@article{arxiv.2603.25836,
title = {Gradient-Informed Training for Low-Resource Multilingual Speech Translation},
author = {Ruiyan Sun and Satoshi Nakamura},
journal= {arXiv preprint arXiv:2603.25836},
year = {2026}
}