Personalized head-related transfer function (HRTF) modeling is essential for immersive spatial audio rendering. However, acquiring high-resolution HRTFs is laborious, as it requires measurements from hundreds or thousands of spatial directions. To alleviate this burden, recent approaches either estimate HRTFs from anatomical data or reconstruct high-resolution HRTFs from a small set of measured directions. Existing methods, however, are often limited to single-dataset training and may fail to generalize to new environments due to dataset-specific measurement biases (e.g., device frequency responses and recording environment setup). In this work, we propose DACD-HRTF, a cross-dataset personalized HRTF upsampling framework based on sparse HRTF measurements. Specifically, DACD-HRTF integrates a dual-branch direction-aware autoencoder to encode sparse measurements into subject-related and dataset-related representations, which are then decoded to estimate personalized HRTFs over high-resolution spatial directions. This design supports joint training across multiple datasets with heterogeneous spatial sampling grids and measurement conditions. To further mitigate measurement biases across datasets, we introduce a contrastive-based disentanglement strategy that encourages the separation of subject-related and dataset-related information, thereby enhancing cross-dataset generalization. Crucially, our framework enables direct reconstruction from sparse measurements without fine-tuning of network parameters at inference, allowing rapid deployment in new environments. Extensive experiments on eight public HRTF datasets demonstrate that DACD-HRTF achieves competitive performance in reconstructing personalized HRTFs for unseen datasets, showing particular effectiveness under practical sparse measurement settings.
Real-time audio stream segmentation is a critical prerequisite for seamless interaction in resource-constrained embodied intelligence (EI). Traditional methods treat segmentation as a passive postprocessing step within automatic speech recognition (ASR) pipelines, leading to accumulated latency and semantic fragmentation. In this paper, we propose PRIMA, a lightweight end-to-end proactive semantic segmentation framework. The architecture comprises three hierarchical modules: a context-aware acoustic frontend for boundary perception, an attentive segmentation encoder for condensing frame-level streams into latent representations, and a multi-stage semantic refiner. To bridge the modality gap, we implement a two-phase progressive distillation strategy. Phase 1 (Global Semantic Following) establishes a macroscopic cross-modal mapping utilizing priors from XLM-RoBERTa and Qwen2.5-Omni. Phase 2 (Recursive Semantic Grounding) employs our Semantic-Aware Cross-modal Alignment (SACA) algorithm to dynamically aggregate acoustic sequences into precise and logically coherent semantic anchors. Experimental results demonstrate that PRIMA achieves a relative reduction of 43.4% in Character Error Rate (CER) compared to traditional ASR pipelines and performs on par with the state-of-the-art semantic voice activity detection (VAD). Furthermore, the framework achieves a real-time factor (RTF) of 0.036 on a server-grade NVIDIA T4 GPU and maintains a robust RTF of 0.3117 on an edge-side Raspberry Pi 5. These results validate PRIMA's efficacy in delivering low-latency, resilient semantic partitioning for real-time edge deployment without relying on the high-performance cloud infrastructure.
Automated Depression Detection (ADD) from speech aims to estimate an individual's depression-related attributes using artificial intelligence applied to spoken signals. However, existing speech-based ADD methods rely on handcrafted acoustic features and high-resource language, which fail to capture clinically grounded speech cues and limit their effectiveness in low-resource multilingual settings. In addition, the integration of these features to overcome individual heterogeneity remains a significant challenge. In this regard, we propose a novel, joint representation learning framework specifically designed to support multilingual ADD. The design of simple yet effective cross-attention mechanisms to fuse the embeddings of speech foundation models with acoustic features offers the ability to enhance contextual information. Notably, this study introduces a new Indic-Bengali dataset on depression speech, collected through standardized and comprehensive elicitation tasks, with diagnostic labels provided by a psychiatrist. Using the proposed framework, extensive subject-independent experiments were conducted on the DAIC-WOZ, EDAIC, Androids corpus, and our Indic-Bengali dataset. We demonstrate that our method outperforms existing state-of-the-art speech-based ADD approaches on classification tasks, while also exhibiting improved cross-corpus transferability in multilingual settings. Finally, we believe that the proposed dataset and framework may support future research on speech-based depression assessment.
Large language models have shown strong generalpurpose capabilities, but they often remain limited in knowledgeintensive reasoning due to insufficient domain-specific knowledge and factual unreliability. Existing methods for integrating knowledge graphs into LLMs either rely on prompt-based knowledge injection, which may lose structural semantics, or use a single adaptation module, which is often insufficient for modeling diverse and context-dependent knowledge. To address these challenges, we propose we propose MoKA, a Mixture of Knowledgeable Adaptations framework for effective knowledge integration and reasoning. MoKA first employs a structural knowledge embedding pre-training module to encode relational semantics from KGs. A knowledgeable adaptation (KA) module then dynamically infuses entity-linked structural knowledge into the LLM through cross-attention interactions. To capture diverse knowledge semantics, MoKA further introduces multiple knowledgeable adaptation experts within Transformer feed-forward blocks and employs an entropy-regularized routing strategy to encourage selective routing and balanced expert utilization. Experiments on CommonsenseQA, SIQA, TriviaQA,WebQuestionsSP, and BIG-Bench Hard show that MoKA improves upon the compared baselines in many evaluated settings across different backbones while maintaining high parameter efficiency. These results support the effectiveness of the combined framework for knowledge-aware LLM reasoning; embedding controls that preserve entity linking do not establish KG semantics as the primary source of gain.
Independent vector analysis (IVA), a fundamental source separation technique, can be extended by incorporating prior knowledge of the sources and their spatial characteristics. When pre-estimated relative transfer functions (RTFs) are available for one or more sources, spatially-regularized IVA (SR-IVA) leverages this information as spatial regularization to enhance IVA. However, SR-IVA suffers from a fundamental scale inconsistency: the optimal scaling of each separation filter differs between the IVA cost function and the spatial regularization term, which calls the validity of the SR-IVA formulation into question. Scale-consistent SR-IVA (SSR-IVA) was proposed to address this limitation. While it retains the IVA cost function for optimizing the separation matrix, it applies spatial regularization only to its scale-consistent part, specifically, the inverse of the RTF matrix, each column of which models the RTF of a source. This formulation effectively resolves the scale inconsistency inherent in SR-IVA. Despite this modeling advantage, the existing optimization algorithm for SSR-IVA suffers from slow convergence, which limits its practical applicability. In this paper, we propose two block majorization-minimization algorithms for SSR-IVA that sequentially update each RTF in a computationally efficient manner. These algorithms can be regarded as extensions of the conventional iterative source steering (ISS) algorithm for IVA and achieve theoretically minimal per-iteration complexity. As a key technique, we introduce an RTF-structure-preserving multiplicative update to ensure that the RTF matrix remains composed of the RTFs throughout the optimization. Experimental results demonstrate that SSR-IVA using the proposed ISS-type algorithms achieves higher source separation accuracy than both SR-IVA and SSR-IVA using the conventional algorithm, while also converging significantly faster than the latter.
