Discourse-Aware Dual-Track Streaming Response for Low-Latency Spoken Dialogue Systems

Achieving human-like responsiveness is a critical yet challenging goal for cascaded spoken dialogue systems. Conventional ASR-LLM-TTS pipelines follow a strictly sequential paradigm, requiring complete transcription and full reasoning before speech synthesis can begin, which results in high response latency. We propose the Discourse-Aware Dual-Track Streaming Response (DDTSR) framework, a low-latency architecture that enables listen-while-thinking and speak-while-thinking. DDTSR is built upon three key mechanisms: (1) connective-guided small-large model synergy, where an auxiliary small model generates minimal-committal discourse connectives while a large model performs knowledge-intensive reasoning in parallel; (2) streaming-based cross-modal collaboration, which dynamically overlaps ASR, LLM inference, and TTS to advance the earliest speakable moment; and (3) curriculum-learning-based discourse continuity enhancement, which maintains coherence and logical consistency between early responses and subsequent reasoning outputs. Experiments on two spoken dialogue benchmarks demonstrate that DDTSR reduces response latency by 19%-51% while preserving discourse quality. Further analysis shows that DDTSR functions as a plug-and-play module compatible with diverse LLM backbones, and remains robust across varying utterance lengths, indicating strong practicality and scalability for real-time spoken interaction.

Deepfake Word Detection by Next-token Prediction using Fine-tuned Whisper

Deepfake speech utterances can be forged by replacing one or more words in a bona fide utterance with semantically different words synthesized by speech generative models. While a dedicated synthetic word detector could be developed, we investigate a cost-effective method that fine-tunes a pre-trained Whisper model to detect synthetic words while transcribing the input utterance via next-token prediction. We further investigate using partially vocoded utterances as the fine-tuning data, thereby reducing the cost of data collection. Our experiments demonstrate that, on in-domain test data, the fine-tuned Whisper yields low synthetic-word detection error rates and transcription error rates. On out-of-domain test data with synthetic words produced by unseen speech generative models, the fine-tuned Whisper remains on par with a dedicated ResNet-based detection model; however, the overall performance degradation calls for strategies to improve its generalization capability.

A Mixture-of-Experts Model for Multimodal Emotion Recognition in Conversations

Emotion Recognition in Conversations (ERC) presents unique challenges, requiring models to capture the temporal flow of multi-turn dialogues and to effectively integrate cues from multiple modalities. We propose Mixture of Speech-Text Experts for Recognition of Emotions (MiSTER-E), a modular Mixture-of-Experts (MoE) framework designed to decouple two core challenges in ERC: modality-specific context modeling and multimodal information fusion. MiSTER-E leverages large language models (LLMs) fine-tuned for both speech and text to provide rich utterance-level embeddings, which are then enhanced through a convolutional-recurrent context modeling layer. The system integrates predictions from three experts-speech-only, text-only, and cross-modal-using a learned gating mechanism that dynamically weighs their outputs. To further encourage consistency and alignment across modalities, we introduce a supervised contrastive loss between paired speech-text representations and a KL-divergence-based regulariza-tion across expert predictions. Importantly, MiSTER-E does not rely on speaker identity at any stage. Experiments on three benchmark datasets-IEMOCAP, MELD, and MOSI-show that our proposal achieves 70.9%, 69.5%, and 87.9% weighted F1-scores respectively, outperforming several baseline speech-text ERC systems. We also provide various ablations to highlight the contributions made in the proposed approach.

Make It Hard to Hear, Easy to Learn: Long-Form Bengali ASR and Speaker Diarization via Extreme Augmentation and Perfect Alignment

Although Automatic Speech Recognition (ASR) in Bengali has seen significant progress, processing long-duration audio and performing robust speaker diarization remain critical research gaps. To address the severe scarcity of joint ASR and diarization resources for this language, we introduce Lipi-Ghor-882, a comprehensive 882-hour multi-speaker Bengali dataset. In this paper, detailing our submission to the DL Sprint 4.0 competition, we systematically evaluate various architectures and approaches for long-form Bengali speech. For ASR, we demonstrate that raw data scaling is ineffective; instead, targeted fine-tuning utilizing perfectly aligned annotations paired with synthetic acoustic degradation (noise and reverberation) emerges as the singular most effective approach. Conversely, for speaker diarization, we observed that global open-source state-of-the-art models (such as Diarizen) performed surprisingly poorly on this complex dataset. Extensive model retraining yielded negligible improvements; instead, strategic, heuristic post-processing of baseline model outputs proved to be the primary driver for increasing accuracy. Ultimately, this work outlines a highly optimized dual pipeline achieving a $\sim$0.019 Real-Time Factor (RTF), establishing a practical, empirically backed benchmark for low-resource, long-form speech processing.

Relating the Neural Representations of Vocalized, Mimed, and Imagined Speech

We investigated the relationship among neural representations of vocalized, mimed, and imagined speech recorded using publicly available stereotactic EEG recordings. Most prior studies have focused on decoding speech responses within each condition separately. Here, instead, we explore how responses across conditions relate by training linear spectrogram reconstruction models for each condition and evaluate their generalization across conditions. We demonstrate that linear decoders trained on one condition generally transfer successfully to others, implying shared speech representations. This commonality was assessed with stimulus-level discriminability by performing a rank-based analysis demonstrating preservation of stimulus-specific structure in both within- and across-conditions. Finally, we compared linear reconstructions to those from a nonlinear neural network. While both exhibited cross-condition transfer, linear models achieve superior stimulus-level discriminability.

Efficient Dialect-Aware Modeling and Conditioning for Low-Resource Taiwanese Hakka Speech Processing

Taiwanese Hakka is a low-resource, endangered language that poses significant challenges for automatic speech recognition (ASR), including high dialectal variability and the presence of two distinct writing systems (Hanzi and Pinyin). Traditional ASR models often encounter difficulties in this context, as they tend to conflate essential linguistic content with dialect-specific variations across both phonological and lexical dimensions. To address these challenges, we propose a unified framework grounded in the Recurrent Neural Network Transducers (RNN-T). Central to our approach is the introduction of dialect-aware modeling strategies designed to disentangle dialectal "style" from linguistic "content", which enhances the model's capacity to learn robust and generalized representations. Additionally, the framework employs parameter-efficient prediction networks to concurrently model ASR (Hanzi and Pinyin). We demonstrate that these tasks create a powerful synergy, wherein the cross-script objective serves as a mutual regularizer to improve the primary ASR tasks. Experiments conducted on the HAT corpus reveal that our model achieves 57.00% and 40.41% relative error rate reduction on Hanzi and Pinyin ASR, respectively. To our knowledge, this is the first systematic investigation into the impact of Hakka dialectal variations on ASR and the first single model capable of jointly addressing these tasks.

Modality Collapse as Mismatched Decoding: Information-Theoretic Limits of Multimodal LLMs

Multimodal LLMs can process speech and images, but they cannot hear a speaker's voice or see an object's texture. We show this is not a failure of encoding: speaker identity, emotion, and visual attributes survive through every LLM layer (3--55$\times$ above chance in linear probes), yet removing 64--71% of modality-specific variance improves decoder loss. The decoder has no learned use for these directions; their presence is noise. We formalize this as a mismatched decoder problem: a decoder trained on text can only extract information along text-aligned directions. Accessible information is bounded by the Generalized Mutual Information (GMI), with degradation scaling with distributional distance and decoder sensitivity. The bound is a property of the decoder's scoring rule, not of any particular architecture; it applies whether non-text inputs arrive through a learned projection, a discrete codebook, or no explicit adapter at all. We validate this across five models spanning speech and vision. A controlled experiment (two Prismatic VLMs differing only in encoder text-alignment) confirms the bottleneck is the decoder's scoring rule, not the encoder or projection. A LoRA intervention demonstrates the fix: training with an emotion objective improves emotion accessibility ($+$7.5%) without affecting other attributes, confirming that the training objective determines what becomes accessible.

Moving Speaker Separation via Parallel Spectral-Spatial Processing

Multi-channel speech separation in dynamic environments is challenging as time-varying spatial and spectral features evolve at different temporal scales. Existing methods typically employ sequential architectures, forcing a single network stream to simultaneously model both feature types, creating an inherent modeling conflict. In this paper, we propose a dual-branch parallel spectral-spatial (PS2) architecture that separately processes spectral and spatial features through parallel streams. The spectral branch uses a bi-directional long short-term memory (BLSTM)-based frequency module, a Mamba-based temporal module, and a self-attention module to model spectral features. The spatial branch employs bi-directional gated recurrent unit (BGRU) networks to process spatial features that encode the evolving geometric relationships between sources and microphones. Features from both branches are integrated through a cross-attention fusion mechanism that adaptively weights their contributions. Experimental results demonstrate that the PS2 outperforms existing state-of-the-art (SOTA) methods by 1.6-2.2 dB in scale-invariant signal-to-distortion ratio (SI-SDR) for moving speaker scenarios, with robust separation quality under different reverberation times (RT60), noise levels, and source movement speeds. Even with fast source movements, the proposed model maintains SI-SDR improvements of over 13 dB. These improvements are consistently observed across multiple datasets, including WHAMR! and our generated WSJ0-Demand-6ch-Move dataset.

TG-ASR: Translation-Guided Learning with Parallel Gated Cross Attention for Low-Resource Automatic Speech Recognition

Low-resource automatic speech recognition (ASR) continues to pose significant challenges, primarily due to the limited availability of transcribed data for numerous languages. While a wealth of spoken content is accessible in television dramas and online videos, Taiwanese Hokkien exemplifies this issue, with transcriptions often being scarce and the majority of available subtitles provided only in Mandarin. To address this deficiency, we introduce TG-ASR for Taiwanese Hokkien drama speech recognition, a translation-guided ASR framework that utilizes multilingual translation embeddings to enhance recognition performance in low-resource environments. The framework is centered around the parallel gated cross-attention (PGCA) mechanism, which adaptively integrates embeddings from various auxiliary languages into the ASR decoder. This mechanism facilitates robust cross-linguistic semantic guidance while ensuring stable optimization and minimizing interference between languages. To support ongoing research initiatives, we present YT-THDC, a 30-hour corpus of Taiwanese Hokkien drama speech with aligned Mandarin subtitles and manually verified Taiwanese Hokkien transcriptions. Comprehensive experiments and analyses identify the auxiliary languages that most effectively enhance ASR performance, achieving a 14.77% relative reduction in character error rate and demonstrating the efficacy of translation-guided learning for underrepresented languages in practical applications.

OmniZip: Learning a Unified and Lightweight Lossless Compressor for Multi-Modal Data

Lossless compression is essential for efficient data storage and transmission. Although learning-based lossless compressors achieve strong results, most of them are designed for a single modality, leading to redundant compressor deployments in multi-modal settings. Designing a unified multi-modal compressor is critical yet challenging, as different data types vary largely in format, dimension, and statistics. Multi-modal large language models offer a promising resolution but remain too complex for practical use. Thus, we propose \textbf{OmniZip}, \textbf{a unified and lightweight lossless compressor for multi-modal data (like image, text, speech, tactile, database, and gene sequence)}. Built on a lightweight backbone, OmniZip incorporates three key components to enable efficient multi-modal lossless compression: a modality-unified tokenizer that reversibly transforms diverse data into tokens, a modality-routing context learning mechanism that enables flexible multi-modal context modeling, and a modality-routing feedforward design that further enhances the model's nonlinear representation flexibility. A reparameterization training strategy is used to enhance model capacity. OmniZip outperforms or matches other state-of-the-art compressors on multiple modalities, achieving 42\%, 57\%, 62\% and 42\%, 53\% higher compression efficiency than gzip on CLIC-M, TouchandGo, enwik9, LibriSpeech, and WikiSQL datasets, respectively. It also supports near real-time inference on resource-constrained edge devices, reaching about 1MB/s on MacBook CPUs and iPhone NPUs. Our code is released at https://github.com/adminasmi/OmniZip-CVPR2026.