arXiv subjects

Hongjin Song

Publications and source records attributed to Hongjin Song.

6 recordsLinked to original sources

arXiv2026

Joint Residual Reweighting for Classifier Free Guidance in Flow-Matching Zero-Shot TTS

Classifier-free guidance (CFG) is widely used in flow-matching-based zero-shot text-to-speech (TTS), where generation is conditioned on text content and a speech prompt. Standard CFG uses a single guidance weight for their joint conditional effect, while branch-selective guidance emphasizes text or speaker conditioning and can introduce a trade-off between text accuracy and speaker similarity. In this paper, we revisit CFG under independently masked conditions and decompose the guidance field into text, speaker, and joint residuals. We show that condition-specific branch differences couple the joint residual with the corresponding text or speaker residual under a shared weight. Trajectory analysis further shows that the joint residual varies over flow time and contains information that cannot be represented by reweighting the text and speaker residuals alone. Based on these observations, we propose joint residual reweighting, which assigns independent weights to the three residuals. Experiments on F5-TTS, CosyVoice2, and GLM-TTS across three evaluation sets show overall improvements in speaker similarity and text accuracy over the default CFG settings without retraining.

eess.AS↗

arXiv2026

Rethinking Length-Based Training: Batch Composition and Loss Normalization in Speech Token Language Models

Short-to-long training is a simple curriculum for speech models, but its gains can be difficult to interpret. In speech token language models, length-based training can change the shuffle policy, batch composition, token retention, and token weights under batch-mean loss. We disentangle these factors through matched comparisons. In the tested settings, short-to-long ordering shows no independent benefit when batch composition and token exposure are fixed. First-epoch grouping lowers perplexity for Mimi under batch-mean loss, but this gain is not observed under token-balanced loss. The cross-tokenizer results are consistent with a link between chunk-length variation and token weighting. This work provides a systematic analysis protocol for studying length-based training in variable-length speech models.

cs.CL↗

arXiv2026

REVE: Efficient Hallucination Correction for Large Audio-Language Models via Reused Encoder States

Large audio-language models may mention acoustic events that are absent from the input. A separate audio event detector can verify these mentions, but doing so requires a second audio encoder and a separate forward pass. We propose Reused Encoder States for Verifying Events (REVE), a lightweight method that uses states already computed by the target model. One readout summarizes class scores across audio frames, while another uses pooled states from four consecutive frame intervals. Class-aware score fusion combines their outputs to verify generated event mentions without encoding the audio again. On AudioSet, REVE removes 92.9% of label-unsupported mentions under a faithful-mention recall constraint. With fewer added parameters and no second audio-encoding pass, REVE achieves a reduction comparable to those of CED-Tiny and CED-Base. Its complete verification latency is about 1/18 of the CED-Base path. Results on controlled DESED mixtures and different target-model architectures further confirm the effectiveness of encoder-state reuse.

cs.SD↗

arXiv2026

Likelihood-Constrained Acoustic Reranking for Training-Free Hallucination Mitigation in LLM-Based ASR

Large language model (LLM)-based automatic speech recognition (ASR) systems achieve strong performance on conventional speech data by leveraging powerful linguistic priors and multilingual capabilities. However, under challenging conditions, these priors can override acoustic evidence, resulting in unintended translation, instruction execution, repetition, or catastrophic deletion. We propose Likelihood-Constrained Acoustic Reranking (LCAR), a training-free decoding method that improves acoustic grounding while preserving support from the base model. At each decoding step, LCAR first retains tokens whose base-model likelihood falls within a margin of the greedy token, then reranks them using an acoustic compatibility score computed from attention-pooled audio embeddings and the existing LM head. By restricting acoustic intervention to plausible, model-supported alternatives, LCAR requires no additional training, external detector, reference transcript, or auxiliary model at inference. We evaluate LCAR on four LLM-based ASR systems using human-audited TTS and open-source speech challenge suites. At $δ=0.60$, LCAR removes 38.8--57.1\% of detector-identified hallucination failures while largely maintaining WER/CER on standard open-source test sets.

eess.AS↗

arXiv2025

Applying Ensemble Models based on Graph Neural Network and Reinforcement Learning for Wind Power Forecasting

Accurately predicting the wind power output of a wind farm across various time scales utilizing Wind Power Forecasting (WPF) is a critical issue in wind power trading and utilization. The WPF problem remains unresolved due to numerous influencing variables, such as wind speed, temperature, latitude, and longitude. Furthermore, achieving high prediction accuracy is crucial for maintaining electric grid stability and ensuring supply security. In this paper, we model all wind turbines within a wind farm as graph nodes in a graph built by their geographical locations. Accordingly, we propose an ensemble model based on graph neural networks and reinforcement learning (EMGRL) for WPF. Our approach includes: (1) applying graph neural networks to capture the time-series data from neighboring wind farms relevant to the target wind farm; (2) establishing a general state embedding that integrates the target wind farm's data with the historical performance of base models on the target wind farm; (3) ensembling and leveraging the advantages of all base models through an actor-critic reinforcement learning framework for WPF.

cs.LG↗

arXiv2024

DST-GTN: Dynamic Spatio-Temporal Graph Transformer Network for Traffic Forecasting

Accurate traffic forecasting is essential for effective urban planning and congestion management. Deep learning (DL) approaches have gained colossal success in traffic forecasting but still face challenges in capturing the intricacies of traffic dynamics. In this paper, we identify and address this challenges by emphasizing that spatial features are inherently dynamic and change over time. A novel in-depth feature representation, called Dynamic Spatio-Temporal (Dyn-ST) features, is introduced, which encapsulates spatial characteristics across varying times. Moreover, a Dynamic Spatio-Temporal Graph Transformer Network (DST-GTN) is proposed by capturing Dyn-ST features and other dynamic adjacency relations between intersections. The DST-GTN can model dynamic ST relationships between nodes accurately and refine the representation of global and local ST characteristics by adopting adaptive weights in low-pass and all-pass filters, enabling the extraction of Dyn-ST features from traffic time-series data. Through numerical experiments on public datasets, the DST-GTN achieves state-of-the-art performance for a range of traffic forecasting tasks and demonstrates enhanced stability.

cs.AI↗