With the increasing adoption of intelligent voice services in electric power customer service centers, identifying potential information risks from large-scale speech interactions has become essential for service compliance and information security management. However, existing speech-based risk identification methods often struggle with implicit semantic relationships, multi-turn contextual dependencies, and recognition errors under noisy acoustic conditions. To address these challenges, this paper proposes a digital speech information risk identification framework integrating Automatic Speech Recognition (ASR), Abstract Meaning Representation (AMR), and multi-granularity graph neural networks. Specifically, a dual-encoder architecture combines hierarchical Transformer-based textual representations with graph neural network-based AMR encoding to extract complementary contextual and structured semantic features. Clause-level and token-level adjacency structures, together with hierarchical self-attention and cross-attention fusion, are employed to capture long-range semantic dependencies and risk-related entity relationships. Furthermore, a Tacotron2–MelGAN-based speech augmentation module is incorporated to support diverse risk-oriented speech scenarios and robustness evaluation. The framework is designed to identify potential sensitive information disclosure, abnormal service statements, procedural violations, and contextual risks in customer service conversations. Comparative experiments involving SRL-BERT, DEP-GNN, and Lin-AMR, together with ablation studies, are designed to evaluate identification performance using Precision, Recall, Macro-F1, and AUPRC under realistic, simulated, and noise-augmented speech conditions.