Recent advancements in AI-synthesized speech have resulted in highly realistic deepfake audio, posing severe threats to authentication systems and digital media trust. Existing detection models struggle to generalize across diverse synthesis methods, especially those involving neural codec-based Audio Language Models (ALMs). In this work, we propose UniTector++, a novel prosody-aware, multi-stream detection architecture that generalizes across vocoder- and codec-based synthesis. UniTector++ incorporates three complementary streams—Whisper-based semantic embeddings, high-level prosodic features, and codec artifact representations—fused through a Multi-Domain Adaptive Graph Attention Fusion (MAGAF) module. Furthermore, an Emotion-Consistency Verification Module (ECVM) reinforces alignment between speech style and prosodic content, and a Universal Adversarial Robustness (UAR) head improves resistance against adversarial attacks. Evaluated on three benchmark datasets—ASVspoof2021, PolyFake, and Codecfake—UniTector++ achieves state-of-the-art performance with average Equal Error Rate (EER) of 0.57% under unseen synthesis scenarios, outperforming competitive baselines by a relative margin of 28%. Our results demonstrate the model’s superior generalization, interpretability, and robustness, offering a significant advancement in universal deepfake speech detection.

Human emotions are intricate and difficult to decipher through various modalities. Current methodologies frequently employ inflexible fusion strategies that do not consider the dynamic and context-sensitive characteristics of emotional expressions in both visual and textual mediums. This paper presents SYMPHONIA (Synchronizing Facial and Textual Modalities for Emotion Understanding), an innovative architecture engineered to capture and amalgamate emotional signals from facial expressions and language, attuned to contextual and modality interactions. There are two parts to SYMPHONIA: a Facial Emotion Branch that uses Vision Transformers and facial landmarks, and a Textual Emotion Branch that uses RoBERTa embeddings and graph-based reasoning. A Dual-Branch Dynamic Attention Mechanism and a Hierarchical Adaptive Fusion Module are used to connect these branches. SYMPHONIA beat the best models on four datasets: IEMOCAP, MELD, CMU-MOSI, and CMU-MOSEI. It got 80.9% accuracy and 80.1% F1-score on IEMOCAP, which was better than Dualgats (74.8%) and EmoCLIP (75.3%). SYMPHONIA got 74.2% accuracy and 73.5% F1-score for MELD. It beat its competitors by getting a 0.86 Pearson correlation on MOSI and a 0.83 on MOSEI for predicting sentiment. Cross-dataset tests showed that SYMPHONIA could generalize, with 66.9% accuracy when trained on IEMOCAP and tested on MELD. This was better than all the baselines. These results show that SYMPHONIA is good at recognizing emotions and analyzing sentiment in different situations, which shows that it can adapt and do well in different settings.