Introduction: This article stems from the doctoral thesis “Characterization and Classification of Seismo-Volcanic Signals Using Artificial Intelligence Techniques and Data Analytics” at the University of Cauca (2022–2026). Problem: Real-time seismic monitoring requires a rapid and accurate classification process for various types of events to assess volcanic activity and thus activate possible early warnings. However, this field faces a fundamental operational bottleneck: in a newly instrumented volcano, monitoring networks record different types of earthquakes with a lower volume of data than required by standard deep learning classifiers, which need hundreds or thousands of labeled samples to achieve adequate performance. Objective: To overcome the data scarcity problem, we developed and evaluated a Multimodal Prototypical Network (MPN) designed to maintain good classification performance even when labeled data are extremely limited. Methodology: Our approach merges CLIP ViT-B/32 visual embeddings, extracted directly from seismic spectrograms, with a 19-dimensional tabular descriptor vector based on the physical characteristics of the signals under study. This multimodal input is processed using a 5-fold ensemble strategy driven by focal loss. The model was rigorously evaluated using the Chiles-Cerro Negro Volcanic Complex (CCNVC) dataset, categorizing three distinct signal classes: Volcanic-Tectonic (VT), Long-Period (LP), and Very-Long-Period (VLP) events. Results: Under full supervision (6,366 labeled samples), the MPN achieved exceptional performance with a 98.73% balanced accuracy and a 99.89% ROC AUC. Crucially, in a few-shot scenario with only 33 labeled waveforms (14 VT, 14 LP, and 5 VLP), the model retained a highly robust 85.27% balanced accuracy and a 95.36% ROC AUC. Furthermore, physical consistency analysis confirmed that the network’s decisions are driven by spectral morphology (amplitude–geometry correlation, r = 0.248) rather than simple amplitude geometry, aligning with the expected physics of volcanic source mechanisms. Conclusion: These findings establish intermodal transfer from large-scale vision-and-language pretraining as a highly viable and physically consistent strategy for volcanic monitoring environments with a scarcity of labeled data. Originality: Instead of relying solely on raw waveform statistics or standard deep learning, this work uniquely combines foundation models (CLIP) with strict tabular constraints based on physical features to solve the specific few-shot labeling problem inherent in operational volcanology. Limitations: While performance with limited labeled data is noteworthy, the absolute lower bound for labeling requirements (e.g., fewer than 5 VLP samples) was not tested. Furthermore, the model’s physical consistency is tied to the specific origin mechanisms of VT, LP, and VLP events in the CCNVC; applying this exact multimodal framework to signals with greater overlap (such as hybrid events or tremors) might require recalibrating the tabular vector based on the physics of those signals.