Jyh-Shing Roger Jang

Xuanjun Chen, Haibin Wu, Chung-Che Wang et al.

Audio-visual synchronization aims to determine whether the mouth movements and speech in the video are synchronized. VocaLiST reaches state-of-the-art performance by incorporating multimodal Transformers to model audio-visual interact information. However, it requires high computing resources, making it impractical for real-world applications. This paper proposed an MTDVocaLiST model, which is trained by our proposed multimodal Transformer distillation (MTD) loss. MTD loss enables MTDVocaLiST model to deeply mimic the cross-attention distribution and value-relation in the Transformer of VocaLiST. Additionally, we harness uncertainty weighting to fully exploit the interaction information across all layers. Our proposed method is effective in two aspects: From the distillation method perspective, MTD loss outperforms other strong distillation baselines. From the distilled model's performance perspective: 1) MTDVocaLiST outperforms similar-size SOTA models, SyncNet, and Perfect Match models by 15.65% and 3.35%; 2) MTDVocaLiST reduces the model size of VocaLiST by 83.52%, yet still maintaining similar performance.

Xuanjun Chen, Haibin Wu, Helen Meng et al.

Audio-visual active speaker detection (AVASD) is well-developed, and now is an indispensable front-end for several multi-modal applications. However, to the best of our knowledge, the adversarial robustness of AVASD models hasn't been investigated, not to mention the effective defense against such attacks. In this paper, we are the first to reveal the vulnerability of AVASD models under audio-only, visual-only, and audio-visual adversarial attacks through extensive experiments. What's more, we also propose a novel audio-visual interaction loss (AVIL) for making attackers difficult to find feasible adversarial examples under an allocated attack budget. The loss aims at pushing the inter-class embeddings to be dispersed, namely non-speech and speech clusters, sufficiently disentangled, and pulling the intra-class embeddings as close as possible to keep them compact. Experimental results show the AVIL outperforms the adversarial training by 33.14 mAP (%) under multi-modal attacks.

Te-Yu Chi, Yu-Meng Tang, Chia-Wen Lu et al.

Our research focuses on solving the zero-shot text classification problem in NLP, with a particular emphasis on innovative self-training strategies. To achieve this objective, we propose a novel self-training strategy that uses labels rather than text for training, significantly reducing the model's training time. Specifically, we use categories from Wikipedia as our training set and leverage the SBERT pre-trained model to establish positive correlations between pairs of categories within the same text, facilitating associative training. For new test datasets, we have improved the original self-training approach, eliminating the need for prior training and testing data from each target dataset. Instead, we adopt Wikipedia as a unified training dataset to better approximate the zero-shot scenario. This modification allows for rapid fine-tuning and inference across different datasets, greatly reducing the time required for self-training. Our experimental results demonstrate that this method can adapt the model to the target dataset within minutes. Compared to other BERT-based transformer models, our approach significantly reduces the amount of training data by training only on labels, not the actual text, and greatly improves training efficiency by utilizing a unified training set. Additionally, our method achieves state-of-the-art results on both the Yahoo Topic and AG News datasets.

Chung-Che Wang, Yu-Chun Lin, Yu-Teng Hsu et al.

This paper proposes to use both audio input and subject information to predict the personalized preference of two audio segments with the same content in different qualities. A siamese network is used to compare the inputs and predict the preference. Several different structures for each side of the siamese network are investigated, and an LDNet with PANNs' CNN6 as the encoder and a multi-layer perceptron block as the decoder outperforms a baseline model using only audio input the most, where the overall accuracy grows from 77.56% to 78.04%. Experimental results also show that using all the subject information, including age, gender, and the specifications of headphones or earphones, is more effective than using only a part of them.

Yu-Chen Lin, Akhilesh Kumar, Norman Chang et al.

We present four main contributions to enhance the performance of Large Language Models (LLMs) in generating domain-specific code: (i) utilizing LLM-based data splitting and data renovation techniques to improve the semantic representation of embeddings' space; (ii) introducing the Chain of Density for Renovation Credibility (CoDRC), driven by LLMs, and the Adaptive Text Renovation (ATR) algorithm for assessing data renovation reliability; (iii) developing the Implicit Knowledge Expansion and Contemplation (IKEC) Prompt technique; and (iv) effectively refactoring existing scripts to generate new and high-quality scripts with LLMs. By using engineering simulation software RedHawk-SC as a case study, we demonstrate the effectiveness of our data pre-processing method for expanding and categorizing scripts. When combined with IKEC, these techniques enhance the Retrieval-Augmented Generation (RAG) method in retrieving more relevant information, ultimately achieving a 73.33% "Percentage of Correct Lines" for code generation problems in MapReduce applications.

Cheng-Yen Li, Xuanjun Chen, Claire Lin et al.

Large Language Models (LLMs) struggle with knowledge-intensive tasks due to hallucinations and fragmented reasoning over dispersed information. While Retrieval-Augmented Generation (RAG) grounds generation in external sources, existing methods often treat evidence as isolated units, failing to reconstruct the logical chains that connect these dots. Inspired by Complementary Learning Systems (CLS), we propose CodaRAG, a framework that evolves retrieval from passive lookup into active associative discovery. CodaRAG operates via a three-stage pipeline: (1) Knowledge Consolidation to unify fragmented extractions into a stable memory substrate; (2) Associative Navigation to traverse the graph via multi-dimensional pathways-semantic, contextualized, and functional-explicitly recovering dispersed evidence chains; and (3) Interference Elimination to prune hyper-associative noise, ensuring a coherent, high-precision reasoning context. On GraphRAG-Bench, CodaRAG achieves absolute gains of 7-10% in retrieval recall and 3-11% in generation accuracy. These results demonstrate CodaRAG's superior ability to systematically robustify associative evidence retrieval for factual, reasoning, and creative tasks.

Claire Lin, Bo-Han Feng, Xuanjun Chen et al.

Retrieval-Augmented Generation (RAG) has emerged as a promising approach for knowledge-intensive tasks. However, few studies have examined RAG for Taiwanese Historical Archives. In this paper, we present an initial study of a RAG pipeline applied to two historical Traditional Chinese datasets, Fort Zeelandia and the Taiwan Provincial Council Gazette, along with their corresponding open-ended query sets. We systematically investigate the effects of query characteristics and metadata integration strategies on retrieval quality, answer generation, and the performance of the overall system. The results show that early-stage metadata integration enhances both retrieval and answer accuracy while also revealing persistent challenges for RAG systems, including hallucinations during generation and difficulties in handling temporal or multi-hop historical queries.

Wenhua Nie, Binhan Luo, Zijie Meng et al.

Large language models can score well on named game-theory benchmarks while failing on the same strategic computation once semantic cues are removed. We show this gap with procedurally generated zero-sum matrix games: a model that recognizes familiar games drops to 34%, 18%, and 2% success on anonymous $2{\times}2$, $3{\times}3$, and $5{\times}5$ payoff matrices. The benchmark separates semantic recall, learned approximate Nash computation, and an output-interface bottleneck that limits scale. Training only on $2{\times}2$ and $3{\times}3$ games, supervised fine-tuning raises unseen $5{\times}5$--$7{\times}7$ success from 2% to 61%, while exploitability-reward training averages 37% with high seed variance. We prove that the exploitability residual is $2$-Lipschitz in payoff perturbations, unlike discontinuous vertex-returning LP equilibrium selectors, explaining why residual training can transfer under payoff shifts even when formatting instability limits mean performance. A dominated-action padding experiment provides causal evidence: trained models solve $3{\times}3$ games embedded in much larger matrices, while random-padded controls fail and dense $12{\times}12$ games remain near failure. Procedural evaluation is therefore necessary for measuring strategic reasoning, and residual rewards expose a real but format-limited route to approximate equilibrium computation.

Wenhua Nie, ZiCheng Zhu, Jianan Wu et al.

Modern LLM APIs often reveal only top-$K$ logit scores and censor the remaining vocabulary. We study the per-position distribution-recovery limits of this access model. For censoring threshold $τ$, the compatible teacher distributions form an identified set whose total-variation diameter is exactly $U_K=(V-K)\exp(τ)/(Z_A+(V-K)\exp(τ))$, where $Z_A$ is the observed partition function. For KL recovery, we give a computable binary-endpoint lower bound and an asymptotically matching small-ambiguity upper bound, with an extension to reference-aware attackers. Experiments on a Qwen3 math-reasoning teacher reveal a layered extraction hierarchy: on-task top-$K$ distillation recovers 12% of private capability, full-logit distillation recovers 56% despite 99% KL closure, and generation-based extraction recovers 96%. Top-$K$ censoring therefore limits per-position distribution recovery but does not by itself prevent capability extraction, separating fidelity from transfer in prompt-only logit distillation.

Wenhua Nie, Jianan Wu, Junlin Liu et al.

Group Relative Policy Optimization (GRPO) is a standard algorithm for reinforcement learning from verifiable rewards, but its group-mean-centered advantage can fail under binary rewards. The failure mode is gradient starvation: when every response in a group is correct or every response is wrong, the centered advantage is exactly zero and the policy receives no learning signal. We prove that the true degeneracy rate always exceeds the i.i.d. Bernoulli prediction by Jensen's inequality, and observe a 0.69 degeneracy rate at group size four in logged Qwen3.5-9B GSM8K training. We then show that the fixed-reference Sign advantage, $A=2r-1$, performs pass@$G$ failure descent by increasing the probability that at least one sample in the group succeeds. On the full GSM8K test set across seven seeds, Sign reaches 73.8% accuracy versus 28.4% for standard normalized group-mean DrGRPO at group size four, a 45.4 point gain with $p<0.0001$. The effect is directionally consistent on Llama-3.1-8B and positive but underpowered on a MATH-500 transfer check. Pass@$k$ analysis indicates that the main benefit is search compression rather than large capacity expansion, aligning the empirical gains with recent RLVR ceiling observations.

Wenhua Nie, Zijie Meng, Kun Zou et al.

Grammar-constrained generation is often combined with local vocabulary masking and speculative decoding, but the resulting sampling law is not the grammar-conditional distribution users usually intend. We show that any speculative decoder with local mask access, Leviathan rejection, and rollback soundness samples from the locally projected distribution $μ^{\mathrm{proj}}$ rather than the grammar-conditional distribution $μ^\star$. This extends the GAD impossibility result to speculative decoding; on Dyck grammars with Qwen3-8B, the total-variation gap can reach 0.996. We identify the future-validity function $Φ_t(y)=\Pr_p[\mathrm{valid\ completion}\mid y]$ as the missing correction statistic. The target distribution is a Doob transform of the base model with $h=Φ$, while local masking corresponds to setting $h$ to one. With exact $Φ$, our oracle decoder FVO-Spec samples exactly from $μ^\star$; with approximate $Φ$, we bound the resulting total-variation error. Because exact future validity is hard for general context-free grammars, we evaluate estimator hierarchies on tractable Dyck and finite JSON languages. OneStep reduces Dyck TV by 14% with under 1% throughput overhead, exact dynamic programming reduces it by 97%, and finite-language correction closes JSON gaps to numerical precision. All fidelity claims are scoped to enumerable grammars and token tries.

Wenhua Nie, Junlin Liu, Jianan Wu et al.

Chain-of-thought reasoning is often treated as a monotone way to improve language-model accuracy by letting a model think longer. We identify a countervailing effect, the coupling tax: when reasoning traces and final answers share one output-token budget, long traces can crowd out the answer they are meant to support. Across GSM8K, MATH-500, and five BIG-Bench Hard tasks with Qwen3 models at three scales, non-thinking mode matches or outperforms thinking mode on GSM8K and MATH-500 at every budget up to 2048 tokens, while harder tasks shift the crossover to larger budgets. We derive a truncation-waste decomposition, $\mathrm{Acc}_{\mathrm{think}}(b)=α_c F_L(b)+α_t(1-F_L(b))$, that predicts this crossover from chain-length and accuracy statistics and explains inverse scaling within the Qwen family. A DeepSeek-R1-Distill-Llama-8B replication shows the same pattern under a different thinking interface. As a mitigation, split-budget generation decouples reasoning and answer budgets; on full MATH-500, IRIS reaches 74.0% accuracy, a strengthened extraction variant reaches 78.8%, and a fixed non-oracle SC+IRIS gate reaches 83.6%. The results show that test-time reasoning should be evaluated as a budget-allocation problem, not only as a question of whether longer traces are available.

Te-Lun Yang, Jyi-Shane Liu, Yuen-Hsien Tseng et al.

This study develops a question-answering system based on Retrieval-Augmented Generation (RAG) using Chinese Wikipedia and Lawbank as retrieval sources. Using TTQA and TMMLU+ as evaluation datasets, the system employs BGE-M3 for dense vector retrieval to obtain highly relevant search results and BGE-reranker to reorder these results based on query relevance. The most pertinent retrieval outcomes serve as reference knowledge for a Large Language Model (LLM), enhancing its ability to answer questions and establishing a knowledge retrieval system grounded in generative AI. The system's effectiveness is assessed through a two-stage evaluation: automatic and assisted performance evaluations. The automatic evaluation calculates accuracy by comparing the model's auto-generated labels with ground truth answers, measuring performance under standardized conditions without human intervention. The assisted performance evaluation involves 20 finance-related multiple-choice questions answered by 20 participants without financial backgrounds. Initially, participants answer independently. Later, they receive system-generated reference information to assist in answering, examining whether the system improves accuracy when assistance is provided. The main contributions of this research are: (1) Enhanced LLM Capability: By integrating BGE-M3 and BGE-reranker, the system retrieves and reorders highly relevant results, reduces hallucinations, and dynamically accesses authorized or public knowledge sources. (2) Improved Data Privacy: A customized RAG architecture enables local operation of the LLM, eliminating the need to send private data to external servers. This approach enhances data security, reduces reliance on commercial services, lowers operational costs, and mitigates privacy risks.

Chih-Hsiang Hsu, Jyh-Shing Roger Jang

Current approaches in 3D human pose estimation primarily focus on regressing 3D joint locations, often neglecting critical physical constraints such as bone length consistency and body symmetry. This work introduces a recurrent neural network architecture designed to capture holistic information across entire video sequences, enabling accurate prediction of bone lengths. To enhance training effectiveness, we propose a novel augmentation strategy using synthetic bone lengths that adhere to physical constraints. Moreover, we present a bone length adjustment method that preserves bone orientations while substituting bone lengths with predicted values. Our results demonstrate that existing 3D human pose estimation models can be significantly enhanced through this adjustment process. Furthermore, we fine-tune human pose estimation models using inferred bone lengths, observing notable improvements. Our bone length prediction model surpasses the previous best results, and our adjustment and fine-tuning method enhance performance across several metrics on the Human3.6M dataset.

Xuanjun Chen, Chia-Yu Hu, Sung-Feng Huang et al.

Rapid advances in singing voice synthesis have increased unauthorized imitation risks, creating an urgent need for better Singing Voice Deepfake (SingFake) Detection, also known as SVDD. Unlike speech, singing contains complex pitch, wide dynamic range, and timbral variations. Conventional 16 kHz-sampled detectors prove inadequate, as they discard vital high-frequency information. This study presents the first systematic analysis of high-resolution (44.1 kHz sampling rate) audio for SVDD. We propose a joint fullband-subband modeling framework: the fullband captures global context, while subband-specific experts isolate fine-grained synthesis artifacts unevenly distributed across the spectrum. Experiments on the WildSVDD dataset demonstrate that high-frequency subbands provide essential complementary cues. Our framework significantly outperforms 16 kHz-sampled models, proving that high-resolution audio and strategic subband integration are critical for robust in-the-wild detection.

Xuanjun Chen, Shih-Peng Cheng, Jiawei Du et al.

Audio-visual temporal deepfake localization under the content-driven partial manipulation remains a highly challenging task. In this scenario, the deepfake regions are usually only spanning a few frames, with the majority of the rest remaining identical to the original. To tackle this, we propose a Hierarchical Boundary Modeling Network (HBMNet), which includes three modules: an Audio-Visual Feature Encoder that extracts discriminative frame-level representations, a Coarse Proposal Generator that predicts candidate boundary regions, and a Fine-grained Probabilities Generator that refines these proposals using bidirectional boundary-content probabilities. From the modality perspective, we enhance audio-visual learning through dedicated encoding and fusion, reinforced by frame-level supervision to boost discriminability. From the temporal perspective, HBMNet integrates multi-scale cues and bidirectional boundary-content relationships. Experiments show that encoding and fusion primarily improve precision, while frame-level supervision boosts recall. Each module (audio-visual fusion, temporal scales, bi-directionality) contributes complementary benefits, collectively enhancing localization performance. HBMNet outperforms BA-TFD and UMMAFormer and shows improved potential scalability with more training data.

Xuanjun Chen, I-Ming Lin, Lin Zhang et al.

Recent attempts at source tracing for codec-based deepfake speech (CodecFake), generated by neural audio codec-based speech generation (CoSG) models, have exhibited suboptimal performance. However, how to train source tracing models using simulated CoSG data while maintaining strong performance on real CoSG-generated audio remains an open challenge. In this paper, we show that models trained solely on codec-resynthesized data tend to overfit to non-speech regions and struggle to generalize to unseen content. To mitigate these challenges, we introduce the Semantic-Acoustic Source Tracing Network (SASTNet), which jointly leverages Whisper for semantic feature encoding and Wav2vec2 with AudioMAE for acoustic feature encoding. Our proposed SASTNet achieves state-of-the-art performance on the CoSG test set of the CodecFake+ dataset, demonstrating its effectiveness for reliable source tracing.

Tun-Chieh Lou, Chung-Che Wang, Jyh-Shing Roger Jang et al.

This paper proposes the use of iterative transfer learning applied to deep learning models for side-channel attacks. Currently, most of the side-channel attack methods train a model for each individual byte, without considering the correlation between bytes. However, since the models' parameters for attacking different bytes may be similar, we can leverage transfer learning, meaning that we first train the model for one of the key bytes, then use the trained model as a pretrained model for the remaining bytes. This technique can be applied iteratively, a process known as iterative transfer learning. Experimental results show that when using thermal or power consumption map images as input, and multilayer perceptron or convolutional neural network as the model, our method improves average performance, especially when the amount of data is insufficient.

Chun-Hsiang Wang, Chung-Che Wang, Jun-You Wang et al.

Music source separation aims to separate polyphonic music into different types of sources. Most existing methods focus on enhancing the quality of separated results by using a larger model structure, rendering them unsuitable for deployment on edge devices. Moreover, these methods may produce low-quality output when the input duration is short, making them impractical for real-time applications. Therefore, the goal of this paper is to enhance a lightweight model, MMDenstNet, to strike a balance between separation quality and latency for real-time applications. Different directions of improvement are explored or proposed in this paper, including complex ideal ratio mask, self-attention, band-merge-split method, and feature look back. Source-to-distortion ratio, real-time factor, and optimal latency are employed to evaluate the performance. To align with our application requirements, the evaluation process in this paper focuses on the separation performance of the accompaniment part. Experimental results demonstrate that our improvement achieves low real-time factor and optimal latency while maintaining acceptable separation quality.

Yen-Lun Liao, Xuanjun Chen, Chung-Che Wang et al.

The countermeasure (CM) model is developed to protect ASV systems from spoof attacks and prevent resulting personal information leakage in Automatic Speaker Verification (ASV) system. Based on practicality and security considerations, the CM model is usually deployed on edge devices, which have more limited computing resources and storage space than cloud-based systems, confining the model size under a limitation. To better trade off the CM model sizes and performance, we proposed an adversarial speaker distillation method, which is an improved version of knowledge distillation method combined with generalized end-to-end (GE2E) pre-training and adversarial fine-tuning. In the evaluation phase of the ASVspoof 2021 Logical Access task, our proposed adversarial speaker distillation ResNetSE (ASD-ResNetSE) model reaches 0.2695 min t-DCF and 3.54% EER. ASD-ResNetSE only used 22.5% of parameters and 19.4% of multiply and accumulate operands of ResNetSE model.

Yu-Hua Chen, Wen-Yi Hsiao, Tsu-Kuang Hsieh et al.

In this paper, we propose a new dataset named EGDB, that con-tains transcriptions of the electric guitar performance of 240 tab-latures rendered with different tones. Moreover, we benchmark theperformance of two well-known transcription models proposed orig-inally for the piano on this dataset, along with a multi-loss Trans-former model that we newly propose. Our evaluation on this datasetand a separate set of real-world recordings demonstrate the influenceof timbre on the accuracy of guitar sheet transcription, the potentialof using multiple losses for Transformers, as well as the room forfurther improvement for this task.

Hsuan-Yu Lin, Xuanjun Chen, Jyh-Shing Roger Jang

Due to the rapid development of deep learning, we can now successfully separate singing voice from mono audio music. However, this separation can only extract human voices from other musical instruments, which is undesirable for karaoke content generation applications that only require the separation of lead singers. For this karaoke application, we need to separate the music containing male and female duets into two vocals, or extract a single lead vocal from the music containing vocal harmony. For this reason, we propose in this article to use a singer separation system, which generates karaoke content for one or two separated lead singers. In particular, we introduced three models for the singer separation task and designed an automatic model selection scheme to distinguish how many lead singers are in the song. We also collected a large enough data set, MIR-SingerSeparation, which has been publicly released to advance the frontier of this research. Our singer separation is most suitable for sentimental ballads and can be directly applied to karaoke content generation. As far as we know, this is the first singer-separation work for real-world karaoke applications.

Szu-Yu Chou, Kai-Hsiang Cheng, Jyh-Shing Roger Jang et al.

In this paper, we introduce a novel attentional similarity module for the problem of few-shot sound recognition. Given a few examples of an unseen sound event, a classifier must be quickly adapted to recognize the new sound event without much fine-tuning. The proposed attentional similarity module can be plugged into any metric-based learning method for few-shot learning, allowing the resulting model to especially match related short sound events. Extensive experiments on two datasets shows that the proposed module consistently improves the performance of five different metric-based learning methods for few-shot sound recognition. The relative improvement ranges from +4.1% to +7.7% for 5-shot 5-way accuracy for the ESC-50 dataset, and from +2.1% to +6.5% for noiseESC-50. Qualitative results demonstrate that our method contributes in particular to the recognition of transient sound events.

Zhe-Cheng Fan, Yen-Lin Lai, Jyh-Shing Roger Jang

Separating two sources from an audio mixture is an important task with many applications. It is a challenging problem since only one signal channel is available for analysis. In this paper, we propose a novel framework for singing voice separation using the generative adversarial network (GAN) with a time-frequency masking function. The mixture spectra is considered to be a distribution and is mapped to the clean spectra which is also considered a distribtution. The approximation of distributions between mixture spectra and clean spectra is performed during the adversarial training process. In contrast with current deep learning approaches for source separation, the parameters of the proposed framework are first initialized in a supervised setting and then optimized by the training procedure of GAN in an unsupervised setting. Experimental results on three datasets (MIR-1K, iKala and DSD100) show that performance can be improved by the proposed framework consisting of conventional networks.