The Anatomy of an ASR Hallucination
研究解决了ASR系统产生无关文本的问题，通过分析两个Conformer-Large模型在不同条件下的表现，发现最终编码阶段是关键，其失败导致输出失去音频基础。
0 citationsRead paper
Selected work
Representative Papers
Recursive State Inference for Linear PASFA
To address the challenge of efficiently and accurately recovering slow feature states from ARMA-type observed data in linear Probabilistic Adaptive Slow Feature Analysis (PASFA), this paper proposes a recursive state estimation algorithm based on Minimum Mean Square Error (MMSE). The method directly models the ARMA dynamics of slow features, bypassing the explicit state-space transformation required by conventional Kalman filtering—thereby preventing distortion and information loss of the original slow features during transformation. As the first direct recursive estimation algorithm specifically designed for linear PASFA, it combines theoretical rigor with computational efficiency. Experimental validation on synthetic data demonstrates that the algorithm achieves high-precision reconstruction of slow features and significantly improves downstream classification accuracy and signal representation quality.
0 citationsRead paper
Odo: Depth-Guided Diffusion for Identity-Preserving Body Reshaping
Existing methods for human shape editing suffer from severe distortions in body proportions, texture warping, and background inconsistency, compounded by the absence of large-scale benchmark datasets. To address these challenges, this paper introduces the first large-scale dataset specifically designed for human shape editing. We further propose an end-to-end deep-guided diffusion model: the UNet backbone is frozen to preserve identity, pose, and clothing consistency, while a SMPL depth-map-driven ControlNet is tightly coupled to enable fine-grained semantic control over shape deformation. This architecture substantially improves geometric fidelity and visual realism. Quantitative evaluation demonstrates a vertex reconstruction error of only 7.5 mm—significantly lower than the baseline (13.6 mm)—and achieves state-of-the-art performance in both target shape alignment and generation quality.
0 citationsRead paper
Recent publications
Latest Papers
The Anatomy of an ASR Hallucination
研究解决了ASR系统产生无关文本的问题，通过分析两个Conformer-Large模型在不同条件下的表现，发现最终编码阶段是关键，其失败导致输出失去音频基础。
0 citationsRead paper
Recursive State Inference for Linear PASFA
To address the challenge of efficiently and accurately recovering slow feature states from ARMA-type observed data in linear Probabilistic Adaptive Slow Feature Analysis (PASFA), this paper proposes a recursive state estimation algorithm based on Minimum Mean Square Error (MMSE). The method directly models the ARMA dynamics of slow features, bypassing the explicit state-space transformation required by conventional Kalman filtering—thereby preventing distortion and information loss of the original slow features during transformation. As the first direct recursive estimation algorithm specifically designed for linear PASFA, it combines theoretical rigor with computational efficiency. Experimental validation on synthetic data demonstrates that the algorithm achieves high-precision reconstruction of slow features and significantly improves downstream classification accuracy and signal representation quality.
0 citationsRead paper
Odo: Depth-Guided Diffusion for Identity-Preserving Body Reshaping
Existing methods for human shape editing suffer from severe distortions in body proportions, texture warping, and background inconsistency, compounded by the absence of large-scale benchmark datasets. To address these challenges, this paper introduces the first large-scale dataset specifically designed for human shape editing. We further propose an end-to-end deep-guided diffusion model: the UNet backbone is frozen to preserve identity, pose, and clothing consistency, while a SMPL depth-map-driven ControlNet is tightly coupled to enable fine-grained semantic control over shape deformation. This architecture substantially improves geometric fidelity and visual realism. Quantitative evaluation demonstrates a vertex reconstruction error of only 7.5 mm—significantly lower than the baseline (13.6 mm)—and achieves state-of-the-art performance in both target shape alignment and generation quality.
0 citationsRead paper