Abstract: Humanoid robots represent a core area of the embodied intelligence industry, and enabling reliable auditory identity perception in complex open environments remains a major challenge. The accuracy of voiceprint feature extraction directly affects the interaction experience of robots and constitutes a key technology. This study provides an in-depth analysis of the Whisper encoder’s capability to extract multi-layer hidden speech features. Through layer-wise contribution analysis, the preferences and redundancies of different encoder layers in acoustic and semantic encoding are revealed. Accordingly, a two-stage adaptive fusion model is proposed. First, a full-layer global weighting mechanism is used to evaluate the contribution of each hidden layer to identity discrimination and to select key acoustic layers that are strongly related to speaker identity. Then, the selected layers are subjected to projection alignment and secondary adaptive weighted fusion to enhance the complementary representation of multi-layer information. Experiments based on the ECAPA-TDNN architecture show that the proposed model outperforms traditional baselines in standard, cross-domain dialect, and complex real-world noisy environments, with a significant reduction in equal error rate (EER). The findings can support the design of highly reliable robotic auditory and interaction systems.