Video anomaly detection (VAD) aims to identify abnormal events based on video-level annotations. In recent years, multimodal data has helped to address the limitations of relying solely on the visual modality in complex scenarios. However, existing methods exploit multimodal information relatively coarsely. Meanwhile, current methods introduce long-range noise when modeling temporal relationships in videos. To address these issues, this paper proposes a triple-modal-guided video anomaly detection method (TMG-VAD). Specifically, we design an audio optimization module (AOM) that uses visual information to model and enhance raw audio signals in the time–frequency domain, producing more discriminative audio features. Furthermore, we introduce an Abnormal Weight Generation Module (AWGM), which employs a multilayer perceptron (MLP) to generate temporal guidance maps from audio features. The temporal guidance maps reflect the anomaly probability for each video segment. In addition, we propose a Video–Text Alignment Module (VTAM) that incorporates the temporal guidance maps into a selective state-space model (SSM) to produce more representative video features while reducing the influence of long-range noise. It also achieves accurate alignment between visual and textual features through contrastive learning. Extensive experiments conducted on the benchmark datasets UCF-Crime and XD-Violence demonstrate the effectiveness of TMG-VAD.