Speaker verification systems often rely on globally pooled channel descriptors or fixed-scale temporal aggregation, which may overlook frame-aligned frequency summaries and stage-dependent temporal context. These limitations become more critical in reduced-data training scenarios. To address them, we propose HiFT, a Hierarchical Frequency-to-Time Recalibration module for efficient speaker embedding learning. HiFT first derives channel-specific, frame-aligned descriptors from frequency-pooled average and max statistics via channel-interleaved fusion, then uses global and stage-specific segment contexts to generate recalibration weights. We train the HiFT-based system on VoxMini, a reduced subset of VoxCeleb2, and evaluate it on the VoxCeleb1 test sets. The results show that, compared with representative baselines, the HiFT-based system achieves a favorable performance-efficiency trade-off under this reduced-data training setting. Further same-backbone comparisons with widely used recalibration modules support the benefit of the proposed frequency-to-time recalibration design.
