Speech spoofing detection is critical for securing Automatic Speaker Verification (ASV) systems in human-machine interaction applications, such as voice assistants, secure telephony, and biometric authentication. Existing countermeasures (CMs) often suffer significant performance degradation under silence-trimmed conditions due to their reliance on non-speech cues, such as background noise or silence duration. This vulnerability is particularly pronounced in real-world scenarios where audio preprocessing (e.g., compression, segmentation) or adversarial attacks (e.g., injecting genuine silence) manipulate silence segments, compromising ASV reliability. To address this, we propose a Wav2vec-Conformer model integrating a novel Multi-Subspace Relative Position Multi-Head Attention mechanism. By leveraging pre-trained Wav2vec 2.0 XLS-R weights, which prioritize phonetic and contextual speech features, and modeling diverse acoustic patterns across multiple subspaces, our approach minimizes dependence on silence cues. Experiments on the silence-trimmed ASVspoof2019LA dataset demonstrate an Equal Error Rate (EER) of 7.257% and a minimum normalized tandem Detection Cost Function (min t-DCF) of 0.1699, achieving relative reductions of 16.6% and 16.1% compared to the state-of-the-art Wav2vec-AASIST baseline (8.703%, 0.2025). Additionally, our model excels on ASVspoof2021LA Hidden set (EER 13.47%, min t-DCF 0.5498), showcasing robust generalization to unseen attacks. These results highlight the model’s effectiveness in silence-unreliable settings, enhancing ASV security for practical human-machine interaction systems.