arXiv subjects

Yuhang Jia

Publications and source records attributed to Yuhang Jia.

At least 19 recordsLinked to original sources

arXiv2026

EchoEdit: Stabilizing Inversion-Free Audio Editing via Optimal Transport Geometry

Text-guided audio editing with pretrained generative models is commonly implemented through inversion or noising. This topology induces a structural trade-off, as stronger edits require deeper corruption of the very rhythm, transients, timbre, and long-range form that should remain unchanged. Here, we introduce EchoEdit, a training-free and inversion-free framework for real-audio editing that directly constructs an editing field by differencing the drifts conditioned on the source and target prompts. This construction avoids explicit source inversion, paired edit data, and test-time optimization, but its stochastic source marginals introduce uncertainty drift, where small random deviations accumulate along the editing trajectory and can move the edited latent away from the audio data manifold. To address this limitation, we further propose EchoEdit+, an optimal-transport-regularized extension that stabilizes the direct editing path by minimizing the transportation cost between edited variables and the source-conditioned audio manifold. The resulting OT coupling contracts the stochastic displacement at noisy states, keeps model queries closer to the training distribution, and preserves structural information while allowing semantic change. Experiments on sound-effect and music editing demonstrate that EchoEdit+ improves target-prompt alignment and source preservation over inversion-based baselines and the unregularized direct editor. Code and dataset will be released.

cs.SD↗

arXiv2026

H-SAGE: Holistic Speaker-Aware Guided Experts for MoE-based Multi-Talker ASR

Multi-talker Automatic Speech Recognition (MTASR) faces significant challenges in accurately transcribing overlapping speech, particularly under complex high-overlap conditions. While recent Mixture-of-Experts (MoE) approaches have shown promise, they typically rely on frame-independent routing that leads to temporal myopia, and depend solely on the downstream ASR objective, which results in implicit and ungrounded representation learning. To address these limitations, we propose Holistic Speaker-Aware Guided Experts (H-SAGE) for MoE-based MTASR. Specifically, we introduce a Speaker-Aware Global Encoder to capture long-term dependencies, supervised by an auxiliary Overlap-Aware Loss that explicitly guides the model to discern acoustic states. Furthermore, we design a Holistic Gating Mechanism to arbitrate expert selection by jointly evaluating global context and local details. Experiments on LibriSpeechMix demonstrate that H-SAGE achieves consistent improvements over strong baselines, particularly in complex scenarios, validating that explicit acoustic guidance effectively enhances expert collaboration. Our code can be found at https://github.com/NKU-HLT/H-SAGE.

cs.SD↗

arXiv2026

CosyEdit: Unlocking End-to-End Speech Editing Capability from Zero-Shot Text-to-Speech Models

Automatic speech editing aims to modify spoken content based on textual instructions, yet traditional cascade systems rely on explicit temporal alignment and complex preprocessing. To address these limitations, we propose CosyEdit, an end-to-end speech editing model adapted from CosyVoice through task-specific post-training and a complementary training paradigm, which internalizes text--speech alignment while ensuring high consistency between the speech before and after editing. Trained on only 250 hours of supervised data from our curated GigaEdit dataset, our 400M-parameter model achieves reliable speech editing performance. Extensive evaluations show that CosyEdit not only outperforms several billion-parameter language model baselines but also approaches state-of-the-art cascade systems. These results show that robust and efficient speech editing can be unlocked from a zero-shot TTS model through post-training, offering a cost-effective end-to-end solution for high-quality speech editing. Code and audio samples are available at https://cjy1018.github.io/CosyEditDemoPage/.

cs.SD↗

arXiv2026

Interpretable Audio Editing Evaluation via Chain-of-Thought Difference-Commonality Reasoning with Multimodal LLMs

Automatic mean opinion score (MOS) prediction serves as a principled alternative to both subjective listening tests and objective metrics, providing scalable and consistent audio evaluation. Inspired by the LLM-as-Judge paradigm, recent multimodal large language models offer strong perceptual modeling and reasoning capabilities, enabling audio quality assessment. In this work, we address the challenging problem of audio editing evaluation and propose the first natural language-based automated evaluation framework built upon Qwen2-Audio. Two caption-based fine-tuning tasks are introduced to enhance multi-audio understanding, together with a designed Chain-of-Thought prompting strategy to encourage structured, step-by-step reasoning. Experiments show that our framework produces interpretable and logically consistent text-based evaluations, aligning closely with human judgments while outperforming existing baselines. The code and demo are available at https://github.com/NKU-HLT/Eval_Reasoning.

cs.SD↗

arXiv2026

UAT: Unified Audio-Text Diffusion for Audio Generation, Editing, and Captioning

Audio generation and audio-to-text understanding remain largely separate, with diffusion models dominating high-fidelity synthesis and autoregressive (AR) language models driving captioning and semantic prediction. Existing unified approaches typically rely on either heterogeneous modules or AR-centric modeling, which can hinder joint optimization and limit acoustic fidelity. We present UAT, to our knowledge, the first diffusion-centric framework that supports unified audio generation, editing, and captioning. UAT couples continuous latent diffusion for audio with masked discrete diffusion for text, enabling bidirectional audio-text modeling within a shared dual-stream backbone. Experiments show that UAT preserves strong audio generation and editing capabilities while achieving competitive captioning performance, demonstrating a favorable balance between acoustic synthesis and semantic prediction. Demo samples are available at https://UAT-demo.github.io.

eess.AS↗

arXiv2026

CosyEdit2: Speech-Editing-Oriented Reinforcement Learning Unlocks Better Zero-Shot TTS

Speech editing and zero-shot Text-to-Speech (TTS) share a similar generative foundation conditioned on speech prompts, yet speech editing demands far stricter local acoustic consistency with surrounding unedited content. While prior work has shown that Supervised Fine-Tuning (SFT) enables TTS models to acquire functional editing capability, this approach remains fundamentally bottlenecked by imperfect paired editing data and coarse-grained optimization signals. To address these limitations, we propose CosyEdit2, a speech editing model built on a two-stage post-training framework that progresses from supervised editing initialization to editing-oriented Group Relative Policy Optimization (GRPO) over target-speech-free data. Extensive experiments demonstrate that CosyEdit2 not only substantially advances speech editing performance, but also unlocks better zero-shot TTS capability, revealing a deeper mutual relationship between the two tasks. Audio samples are available at https://cjy1018.github.io/CosyEdit2.

cs.SD↗

arXiv2026

Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning

Post-training has become central to turning pretrained large language models (LLMs) into aligned, capable, and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objectives rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary regimes: off-policy learning on externally supplied trajectories and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles -- effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions -- together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes useful behavior across stages and model transitions. Under this view, SFT may serve either support expansion or policy reshaping; preference optimization is usually off-policy reshaping, though online variants move closer to learner-generated states. On-policy RL often improves behavior on learner-generated states, but stronger guidance can also make hard-to-reach reasoning paths reachable. Distillation is often better understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress increasingly depends on coordinated systems design rather than any single dominant objective.

cs.CL↗

arXiv2026

GLAD: Global-Local Aware Dynamic Mixture-of-Experts for Multi-Talker ASR

End-to-end multi-talker automatic speech recognition (MTASR) faces significant challenges in accurately transcribing overlapping speech. A critical bottleneck is that speaker-specific acoustic characteristics, which are essential for distinguishing overlapping speech, are often diluted in deep network layers. To address this, we propose the Global-Local Aware Dynamic Mixture-of-Experts (GLAD) architecture. GLAD introduces a novel routing mechanism that dynamically fuses speaker-aware global context with fine-grained local acoustic details to adaptively guide expert selection. Experiments on the LibriSpeechMix and CH109 datasets demonstrate that GLAD significantly outperforms existing Serialized Output Training (SOT)-based MTASR approaches, exhibiting exceptional robustness in challenging, high-overlap scenarios. To the best of our knowledge, this is the first work to apply a global-local fusion MoE strategy to MTASR.

cs.SD↗

arXiv2026

Towards Automatic Evaluation and High-Quality Pseudo-Parallel Dataset Construction for Audio Editing: A Human-in-the-Loop Method

Audio editing aims to manipulate audio content based on textual descriptions, supporting tasks such as adding, removing, or replacing audio events. Despite recent progress, the lack of high-quality benchmark datasets and comprehensive evaluation metrics remains a major challenge for both assessing audio editing quality and improving the task itself. In this work, we propose a novel approach for audio editing task by incorporating expert knowledge into both the evaluation and dataset construction processes: 1) First, we establish AuditScore, the first comprehensive dataset for subjective evaluation of audio editing, consisting of over 6,300 edited samples generated from 7 representative audio editing frameworks and 23 system configurations. Each sample is annotated by professional raters on three key aspects of audio editing quality: overall Quality, Relevance to editing intent, and Faithfulness to original features. 2) Based on this dataset, we systematically propose AuditEval, a family of automatic MOS-style evaluators tailored for audio editing, covering both SSL-based and LLM-based approaches. It addresses the lack of effective objective metrics and the prohibitive cost of subjective evaluation in this field. 3) We further leverage AuditEval to evaluate and filter a large amount of synthetically mixed editing pairs, mining a high-quality pseudo-parallel subset by selecting the most plausible samples. Comprehensive experiments validate that our expert-informed filtering strategy effectively yields higher-quality data, while also exposing the limitations of traditional objective metrics and the advantages of AuditEval. The dataset, codes and tools can be found at: https://github.com/NKU-HLT/AuditEval.

cs.SD↗

arXiv2026

DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding

Autoregressive (AR) large audio language models (LALMs) such as Qwen-2.5-Omni have achieved strong performance on audio understanding and interaction, but scaling them remains costly in data and computation, and strictly sequential decoding limits inference efficiency. Diffusion large language models (dLLMs) have recently been shown to make effective use of limited training data, and prior work on DIFFA indicates that replacing an AR backbone with a diffusion counterpart can substantially improve audio understanding under matched settings, albeit at a proof-of-concept scale without large-scale instruction tuning, preference alignment, or practical decoding schemes. We introduce DIFFA-2, a practical diffusion-based LALM for general audio understanding. DIFFA-2 upgrades the speech encoder, employs dual semantic and acoustic adapters, and is trained with a four-stage curriculum that combines semantic and acoustic alignment, large-scale supervised fine-tuning, and variance-reduced preference optimization, using only fully open-source corpora. Experiments on MMSU, MMAU, and MMAR show that DIFFA-2 consistently improves over DIFFA and is competitive to strong AR LALMs under practical training budgets, supporting diffusion-based modeling is a viable backbone for large-scale audio understanding. Our code is available at https://github.com/NKU-HLT/DIFFA.git.

cs.SD↗

arXiv2026

AudioEval: Automatic Dual-Perspective and Multi-Dimensional Evaluation of Text-to-Audio-Generation

Text-to-audio (TTA) generation is advancing rapidly, but evaluation remains challenging because human listening studies are expensive and existing automatic metrics capture only limited aspects of perceptual quality. We introduce AudioEval, a large-scale TTA evaluation dataset with 4,200 generated audio samples (11.7 hours) from 24 systems and 126,000 ratings collected from both experts and non-experts across five dimensions: enjoyment, usefulness, complexity, quality, and text alignment. Using AudioEval, we benchmark diverse automatic evaluators to compare perspective- and dimension-level differences across model families. We also propose Qwen-DisQA as a strong reference baseline: it jointly processes prompts and generated audio to predict multi-dimensional ratings for both annotator groups, modeling rater disagreement via distributional prediction and achieving strong performance. We will release AudioEval to support future research in TTA evaluation.

cs.SD↗

arXiv2026

Reflecting Twice before Speaking with Empathy: Self-Reflective Alternating Inference for Empathy-Aware End-to-End Spoken Dialogue

End-to-end Spoken Language Models (SLMs) hold great potential for paralinguistic perception, and numerous studies have aimed to enhance their capabilities, particularly for empathetic dialogue. However, current approaches largely depend on rigid supervised signals, such as ground-truth response in supervised fine-tuning or preference scores in reinforcement learning. Such reliance is fundamentally limited for modeling complex empathy, as there is no single "correct" response and a simple numerical score cannot fully capture the nuances of emotional expression or the appropriateness of empathetic behavior. To address these limitations, we sequentially introduce EmpathyEval, a descriptive natural-language-based evaluation model for assessing empathetic quality in spoken dialogues. Building upon EmpathyEval, we propose ReEmpathy, an end-to-end SLM that enhances empathetic dialogue through a novel Empathetic Self-Reflective Alternating Inference mechanism, which interleaves spoken response generation with free-form, empathy-related reflective reasoning. Extensive experiments demonstrate that ReEmpathy substantially improves empathy-sensitive spoken dialogue by enabling reflective reasoning, offering a promising approach toward more emotionally intelligent and empathy-aware human-computer interactions.

cs.CL↗

arXiv2026

From Contrast to Commonality: Audio Commonality Captioning for Enhanced Audio-Text Cross-modal Understanding in Multimodal LLMs

Audio Captioning (AC) plays a pivotal role in enhancing audio-text cross-modal understanding during the pretraining and finetuning of Multimodal LLMs (MLLMs). To strengthen this alignment, recent works propose Audio Difference Captioning (ADC), which takes multiple audio inputs and encourages the model to describe their differences, thereby promoting fine-grained discrimination. However, despite its effectiveness, ADC introduces a semantic gap between input audios-often rich in diverse events-and the brief, difference-focused short caption. This deviation from AC-style task causes a mismatch with the pretraining objective, leading to catastrophic forgetting. To address this, we propose Audio Commonality Captioning (ACC), a comparably challenging but gentler alternative that guides the model to capture shared semantics across audio clips rather than detailed differences. Experiments show that ACC not only improves audio-text understanding on captioning benchmarks but also better preserves general capabilities across diverse speech and music tasks, confirming its ability to enable more robust cross-modal understanding and achieve a better balance between generalization and task-specific performance in MLLMs.

cs.SD↗

arXiv2025

Cross-Modal Knowledge Distillation for Speech Large Language Models

In this work, we present the first systematic evaluation of catastrophic forgetting and modality inequivalence in speech large language models, showing that introducing speech capabilities can degrade knowledge and reasoning even when inputs remain textual, and performance further decreases with spoken queries. To address these challenges, we propose a cross-modal knowledge distillation framework that leverages both text-to-text and speech-to-text channels to transfer knowledge from a text-based teacher model to a speech LLM. Extensive experiments on dialogue and audio understanding tasks validate the effectiveness of our approach in preserving textual knowledge, improving cross-modal alignment, and enhancing reasoning in speech-based interactions.

cs.CL↗

arXiv2025

TTA-Bench: A Comprehensive Benchmark for Evaluating Text-to-Audio Models

Text-to-Audio (TTA) generation has made rapid progress, but current evaluation methods remain narrow, focusing mainly on perceptual quality while overlooking robustness, generalization, and ethical concerns. We present TTA-Bench, a comprehensive benchmark for evaluating TTA models across functional performance, reliability, and social responsibility. It covers seven dimensions including accuracy, robustness, fairness, and toxicity, and includes 2,999 diverse prompts generated through automated and manual methods. We introduce a unified evaluation protocol that combines objective metrics with over 118,000 human annotations from both experts and general users. Ten state-of-the-art models are benchmarked under this framework, offering detailed insights into their strengths and limitations. TTA-Bench establishes a new standard for holistic and responsible evaluation of TTA systems. The dataset and evaluation tools are open-sourced at https://nku-hlt.github.io/tta-bench/.

cs.SD↗

arXiv2025

Enhancing Gradient-based Discrete Sampling via Parallel Tempering

While gradient-based discrete samplers are effective in sampling from complex distributions, they are susceptible to getting trapped in local minima, particularly in high-dimensional, multimodal discrete distributions, owing to the discontinuities inherent in these landscapes. To circumvent this issue, we combine parallel tempering, also known as replica exchange, with the discrete Langevin proposal and develop the Parallel Tempering enhanced Discrete Langevin Proposal (PTDLP), which are simulated at a series of temperatures. Significant energy differences prompt sample swaps, which are governed by a Metropolis criterion specifically designed for discrete sampling to ensure detailed balance is maintained. Additionally, we introduce an automatic scheme to determine the optimal temperature schedule and the number of chains, ensuring adaptability across diverse tasks with minimal tuning. Theoretically, we establish that our algorithm converges non-asymptotically to the target energy and exhibits faster mixing compared to a single chain. Empirical results further emphasize the superiority of our method in sampling from complex, multimodal discrete distributions, including synthetic problems, restricted Boltzmann machines, and deep energy-based models.

stat.ML↗

arXiv2025

Chinese-LiPS: A Chinese audio-visual speech recognition dataset with Lip-reading and Presentation Slides

Incorporating visual modalities to assist Automatic Speech Recognition (ASR) tasks has led to significant improvements. However, existing Audio-Visual Speech Recognition (AVSR) datasets and methods typically rely solely on lip-reading information or speaking contextual video, neglecting the potential of combining these different valuable visual cues within the speaking context. In this paper, we release a multimodal Chinese AVSR dataset, Chinese-LiPS, comprising 100 hours of speech, video, and corresponding manual transcription, with the visual modality encompassing both lip-reading information and the presentation slides used by the speaker. Based on Chinese-LiPS, we develop a simple yet effective pipeline, LiPS-AVSR, which leverages both lip-reading and presentation slide information as visual modalities for AVSR tasks. Experiments show that lip-reading and presentation slide information improve ASR performance by approximately 8\% and 25\%, respectively, with a combined performance improvement of about 35\%. The dataset is available at https://kiri0824.github.io/Chinese-LiPS/

cs.MM↗

arXiv2025

Enhancing Neural Adaptive Wireless Video Streaming via Lower-Layer Information Exposure and Online Tuning

Deep reinforcement learning (DRL) demonstrates its promising potential in the realm of adaptive video streaming and has recently received increasing attention. However, existing DRL-based methods for adaptive video streaming use only application (APP) layer information, adopt heuristic training methods, and train generalized neural networks with pre-collected data. This paper aims to boost the quality of experience (QoE) of adaptive wireless video streaming by using lower-layer information, deriving a rigorous training method, and adopting online tuning with real-time data. First, we formulate a more comprehensive and accurate adaptive wireless video streaming problem as an infinite stage discounted Markov decision process (MDP) problem by additionally incorporating past and lower-layer information, allowing a flexible tradeoff between QoE and costs for obtaining system information and solving the problem. In the offline scenario (only with pre-collected data), we propose an enhanced asynchronous advantage actor-critic (eA3C) method by jointly optimizing the parameters of parameterized policy and value function. Specifically, we build an eA3C network consisting of a policy network and a value network that can utilize cross-layer, past, and current information and jointly train the eA3C network using pre-collected samples. In the online scenario (with additional real-time data), we propose two continual learning-based online tuning methods for designing better policies for a specific user with different QoE and training time tradeoffs. Finally, experimental results show that the proposed offline policy can improve the QoE by 6.8~14.4% compared to the state-of-arts in the offline scenario, and the proposed online policies can further achieve 6~28% gains in QoE over the proposed offline policy in the online scenario.

cs.MM↗