A Data-Oriented Conceptual Model and Hybrid TDNN-BiLSTM Framework for Context-Aware Speaker Verification in Smart Environments

초록

Smart environments and ambient assisted living systems increasingly rely on speaker verification as an unobtrusive mechanism for user authentication. However, requirements for such systems are often captured only in language-based conceptual models, which suffer from ambiguity, inconsistency, and weak linkage to data-intensive AI implementations. This gap makes it difficult to systematically conceptualize, manage, and build big-data-driven, context-aware speaker verification services for future cyber societies. In this paper, we propose a data oriented conceptual model for context-aware speaker verification that addresses key issues of language-based conceptual modeling. As a concrete instantiation of the proposed conceptual model, we designed a hybrid TDNN-BiLSTM speaker embedding framework that integrates Multi-Head Attention (MHA) pooling and an Additive Angular Margin (AAM) softmax (ArcFace) objective. TDNN layers efficiently capture local acoustic patterns, such as formant transitions and phoneme-level cues, whereas the BiLSTM module models long-range sequential dependencies, such as speaking rhythm and prosody. The MHA pooling layer aggregates frame-level features into fixed-dimensional embeddings, while the AAM-softmax objective optimizes angular margins between speakers in the embedding space, enhancing intra-speaker compactness and inter-speaker separation for open-set verification. Experiments on the VoxCeleb1 test set demonstrate that the proposed implementation achieves an Equal Error Rate (EER) of 0.97% and a minimum Detection Cost Function (minDCF) of 0.0717, indicating that the instantiated framework can deliver highly discriminative embeddings within the proposed conceptual design. These results illustrate how a well-defined conceptual model can effectively guide the development of robust, AI-based speaker verification services for sensor-rich smart environments in the physical world and the future cyber society. © 2026 IEEE.

키워드