For the 1.04-point WER gap, I would use a paired speaker-block bootstrap: sample the 38 speakers with replacement, carry all 2–11 clips of each sampled speaker into both systems, and recompute total word errors / total reference words before taking the difference. That preserves your length-weighted metric and avoids treating repeated clips from one speaker as independent observations. The resulting interval would describe sensitivity to the speaker mix in this subset, without implying national representativeness. The ASR-specific rationale is in Liu & Peng: https://www.isca-archive.org/interspeech_2020/liu20c_interspeech.pdf