Alzheimer’s Disease (AD) is a neurodegenerative disorder that severely affects cognitive function. Early detection is critical for slowing disease progression and improving patients’ quality of life. A typical symptom of AD is the degradation of language abilities, making speech analysis a promising non-invasive and low-cost detection method. However, existing methods mainly focus on static feature extraction, ignoring the temporal information in speech signals. This study proposes a novel method for AD detection method based on shifted windows temporal analysis of speech. The shifted windows technique segments speech paralinguistic word tags into short temporal sequences. Temporal features across different windows are then combined to extract global contextual information. A path signature algorithm is introduced to transform long sequence features into fixed-size vectors, thereby addressing the challenge of long-range dependencies. Additionally, a weighted fusion downsampling method is employed to address batch consistency and sequence alignment issues, thereby enhancing model performance. We conducted experiments on the publicly available Pitt and NCMMSC datasets to validate the effectiveness of the proposed method. The results demonstrate significant improvements in AD detection accuracy, achieving 87.1% and 83.15% on the respective datasets. The proposed method fully leverages temporal information in speech and overcomes the limitations of long sequence dependencies. It offers a new and effective approach for early AD diagnosis, with significant clinical relevance and research implications.