Parameter-matched comparison with LoRA-XS

Conduct systematic parameter-matched comparisons between the FCCA-based methods for multilingual low-resource automatic speech recognition and LoRA-XS to assess their relative parameter-efficiency and recognition performance.

Background

The paper evaluates standard FCCA, AC-FCCA, and AR-FCCA against conventional LoRA baselines under several parameter budgets, using Whisper medium and Qwen3-ASR on multilingual low-resource speech-recognition tasks. However, it does not include systematic comparisons with LoRA-XS, a closely related ultra-low-parameter method that uses singular directions from pretrained weight spectra rather than task-informed Fisher-whitened statistics. Such comparisons would clarify whether the reported advantages of FCCA and its extensions persist against this specifically relevant baseline under matched trainable-parameter budgets.