Jen-Tzung Chien
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Jen-Tzung Chien has a total of 172 co-authors. Here are the top co-authors by number of shared publications:
- Man-Wai Mak
Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong, SAR, China
14 shared publications Chuang-Hua Chueh
Department of Electrical and Computer Engineering, National Chiao Tung University, Hsinchu, Taiwan 30010, ROC
13 shared publicationsHsin-Lung Hsieh
Department of Electrical and Computer Engineering, National Chiao Tung University, Taiwan 30010, Republic of China
10 shared publicationsShinji Watanabe
10 shared publicationsMeng-Sung Wu
Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan 70101, ROC
9 shared publicationsChih-Hsien Huang
Dept. of Comput. Sci. & Inf. Eng., Nat. Cheng Kung Univ., Tainan, Taiwan
7 shared publications
Latest Publications
- Open Access
- 24 August 2026
- Published by Emerald Publishing in APSIPA Transactions on Signal and Information Processing
- Vol. 15  (1) , 397-427
Enhancing instance discrimination for self-supervised adversarial training
A learning machine is prone to be attacked by adversarial examples contaminated by perturbations. Adversarial training (AT) is developed to build a defense model against adversarial attacks. Most of previous works were conducted in supervised setting. To release the demand of data labeling, self-supervised learning is required to carry out the self-supervised AT. This study presents a contrastive view in AT where a self-supervised AT is developed by enhancing instance discrimination based on contrastive learning (CL). The connection between AT and self-supervised learning is used to consolidate self-supervised AT by merging a classification perspective in contrastive objective. This study analyzes two potential sources of mismatch in existing CL and AT pipelines, namely, the implicit class definition in instance discrimination and the distributional discrepancy induced by different augmentation strategies. Based on this analysis, this study develops a self-supervised AT framework that integrates a classification perspective into CL. A self-supervised AT with a contrastive view is performed to calculate the perturbed image embeddings to train a robust classifier. The contrastive objective is also reformulated to train a self-supervised sentence encoder as the defense model. The experiments on image and text representations for adversarial robustness show that this method improves the previous unsupervised methods, and even surpasses some other supervised methods.
- 24 August 2026
- Published by Emerald Publishing in APSIPA Transactions on Signal and Information Processing
- Vol. 15  (1) , 397-427
Open AccessEnhancing instance discrimination for self-supervised adversarial training
A learning machine is prone to be attacked by adversarial examples contaminated by perturbations. Adversarial training (AT) is developed to build a defense model against adversarial attacks. Most of previous works were conducted in supervised setting. To release the demand of data labeling, self-supervised learning is required to carry out the self-supervised AT. This study presents a contrastive view in AT where a self-supervised AT is developed by enhancing instance discrimination based on contrastive learning (CL). The connection between AT and self-supervised learning is used to consolidate self-supervised AT by merging a classification perspective in contrastive objective. This study analyzes two potential sources of mismatch in existing CL and AT pipelines, namely, the implicit class definition in instance discrimination and the distributional discrepancy induced by different augmentation strategies. Based on this analysis, this study develops a self-supervised AT framework that integrates a classification perspective into CL. A self-supervised AT with a contrastive view is performed to calculate the perturbed image embeddings to train a robust classifier. The contrastive objective is also reformulated to train a self-supervised sentence encoder as the defense model. The experiments on image and text representations for adversarial robustness show that this method improves the previous unsupervised methods, and even surpasses some other supervised methods.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48  (7) , 8331-8343
Outlier-Aware Contrastive Learning
Contrastive learning aims to learn an embedding space with sample discrimination where similar samples attract together while dissimilar samples repulse apart. However, the issue of sampling bias likely happens and degrades the classification performance when a contrast model is trained with the leakage caused by similar samples but from different classes or dissimilar samples from the same class. Out-of-distribution (OOD) detection provides a meaningful scheme to detect and mask those false negative samples for debiasing in an outlier-aware contrastive loss for high-fidelity contrastive learning. Sample debiasing is feasible to reduce the upper bound of contrastive loss. Also, the previous OOD detector was trained from auxiliary collection of OOD samples. In real world, the prior knowledge of OOD samples is commonly unavailable. This study presents new outlier-aware detection and contrast models through generation and augmentation of those samples near the boundary between in-distribution (ID) and OOD. These synthesized samples are located right outside ID, and their Gaussian embeddings sufficiently reflect OOD behaviors. An OOD detector is learned by using ID samples and synthesized OOD samples with the learning objective towards contrastive OOD detection and debiased contrast model. The experiments are conducted to illustrate the merit of the proposed outlier-aware contrastive learning.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48  (7) , 8331-8343
Outlier-Aware Contrastive Learning
Contrastive learning aims to learn an embedding space with sample discrimination where similar samples attract together while dissimilar samples repulse apart. However, the issue of sampling bias likely happens and degrades the classification performance when a contrast model is trained with the leakage caused by similar samples but from different classes or dissimilar samples from the same class. Out-of-distribution (OOD) detection provides a meaningful scheme to detect and mask those false negative samples for debiasing in an outlier-aware contrastive loss for high-fidelity contrastive learning. Sample debiasing is feasible to reduce the upper bound of contrastive loss. Also, the previous OOD detector was trained from auxiliary collection of OOD samples. In real world, the prior knowledge of OOD samples is commonly unavailable. This study presents new outlier-aware detection and contrast models through generation and augmentation of those samples near the boundary between in-distribution (ID) and OOD. These synthesized samples are located right outside ID, and their Gaussian embeddings sufficiently reflect OOD behaviors. An OOD detector is learned by using ID samples and synthesized OOD samples with the learning objective towards contrastive OOD detection and debiased contrast model. The experiments are conducted to illustrate the merit of the proposed outlier-aware contrastive learning.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 2788-2800
Trilevel Supervised, Unsupervised, and Distilled Learning for Speech Recognition
In this paper, we propose TL-SUD, a trilevel learning framework that unifies supervised, unsupervised and distillation objectives for automatic speech recognition. Unlike two-stage approaches that perform unsupervised pre-training followed by joint fine-tuning with supervised and knowledge distillation objectives, TL-SUD jointly optimizes all three loss functions within a unified process. This enables the acoustic model to learn representations that combine the generalization capability from unlabeled data, the task-specific alignment from labeled data, and the knowledge distillation from a teacher model. To solve the trilevel optimization problem, we sequentially apply penalty-based bilevel gradient descent, first collapsing the lower and middle-levels into a single-level problem and then solving the resulting bilevel structure involving the upper-level. The effectiveness of the proposed framework is evaluated on the LibriSpeech and English Common Voice datasets. We adopt FastConformer as the model backbone, with a supervised connectionist temporal classification loss, an unsupervised contrastive loss, and a feature-based distillation loss. The results show that TL-SUD outperforms both the commonly adopted two-stage pre-training and fine-tuning approach and the weighted-sum methods of the three objectives.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 2788-2800
Trilevel Supervised, Unsupervised, and Distilled Learning for Speech Recognition
In this paper, we propose TL-SUD, a trilevel learning framework that unifies supervised, unsupervised and distillation objectives for automatic speech recognition. Unlike two-stage approaches that perform unsupervised pre-training followed by joint fine-tuning with supervised and knowledge distillation objectives, TL-SUD jointly optimizes all three loss functions within a unified process. This enables the acoustic model to learn representations that combine the generalization capability from unlabeled data, the task-specific alignment from labeled data, and the knowledge distillation from a teacher model. To solve the trilevel optimization problem, we sequentially apply penalty-based bilevel gradient descent, first collapsing the lower and middle-levels into a single-level problem and then solving the resulting bilevel structure involving the upper-level. The effectiveness of the proposed framework is evaluated on the LibriSpeech and English Common Voice datasets. We adopt FastConformer as the model backbone, with a supervised connectionist temporal classification loss, an unsupervised contrastive loss, and a feature-based distillation loss. The results show that TL-SUD outperforms both the commonly adopted two-stage pre-training and fine-tuning approach and the weighted-sum methods of the three objectives.
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. PP  (01628828) , 1-12
Contrastive Mixture Diffusion Models
Continuous diffusion model is feasible to perform smoothed transitions for flexible text generation in a continuous latent space. A challenging issue in continuous diffusion is to implement the denoising process and manipulate the generated tokens to reflect physical phenomenon in diffusion steps towards a contextually and semantically meaningful sentence. Basically, a high-quality text generation can be fulfilled by an easy-first strategy, which prioritizes the generation of high-frequency simple or general tokens in the beginning and then low-frequency complex or specific tokens in the end via mask language model. This paper introduces the mask noise as an absorbing state and develops a new continuous-discrete denoising process in a diffusion model. Each diffusion step is performed by generating either a continuous word embedding or a discrete mask token where the latent transition is modeled by a Gaussian-Dirac mixture distribution. The easy-first text generation is then implemented and strengthened via a contrastive loss to disentangle the generation between simple and complex tokens. A contrastive mixture diffusion model is accordingly exploited by minimizing a variational bound of negative log likelihood, which is regularized by aligning the denoising network posterior with the Gaussian-Dirac posterior in each diffusion transition based on an approximate Kullback-Leibler divergence. The experiments on text paraphrasing and other tasks demonstrate the effectiveness and efficiency of sentence generation by using the proposed method where the easy-first strategy in generation behavior is illustrated.
- 23 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. PP  (01628828) , 1-12
Contrastive Mixture Diffusion Models
Continuous diffusion model is feasible to perform smoothed transitions for flexible text generation in a continuous latent space. A challenging issue in continuous diffusion is to implement the denoising process and manipulate the generated tokens to reflect physical phenomenon in diffusion steps towards a contextually and semantically meaningful sentence. Basically, a high-quality text generation can be fulfilled by an easy-first strategy, which prioritizes the generation of high-frequency simple or general tokens in the beginning and then low-frequency complex or specific tokens in the end via mask language model. This paper introduces the mask noise as an absorbing state and develops a new continuous-discrete denoising process in a diffusion model. Each diffusion step is performed by generating either a continuous word embedding or a discrete mask token where the latent transition is modeled by a Gaussian-Dirac mixture distribution. The easy-first text generation is then implemented and strengthened via a contrastive loss to disentangle the generation between simple and complex tokens. A contrastive mixture diffusion model is accordingly exploited by minimizing a variational bound of negative log likelihood, which is regularized by aligning the denoising network posterior with the Gaussian-Dirac posterior in each diffusion transition based on an approximate Kullback-Leibler divergence. The experiments on text paraphrasing and other tasks demonstrate the effectiveness and efficiency of sentence generation by using the proposed method where the easy-first strategy in generation behavior is illustrated.
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 1594-1607
Dynamic Diffusion Programming and Classification
Connectionist temporal classification (CTC) has been achieving remarkable performance in utilization of neural networks to recognize sequence data like speech signal due to its simplicity, efficiency and stability. The challenging issues in CTC include assuming the conditional temporal independence and tackling the balance between accuracy and efficiency. This paper deals with these two issues by incorporating a probabilistic diffusion model into a non-autoregressive CTC for a new end-to-end automatic speech recognition (ASR). By leveraging the iterative denoising neural network for temporal classification, this method consolidates the dependency modeling in token alignments while maintaining the inference efficiency. In particular, a dynamic diffusion programming algorithm is proposed to implement an efficient computation for optimization where the diffusion objective is calculated over different alignments. Additionally, this study exploits a fast sampling strategy in reverse process, which reduces the latency by skipping those low-confidence diffusion steps. The proposed methods considerably mitigate the computation complexities in diffusion objective and sampling procedure for a discrete-diffusion ASR. The experiments on LibriSpeech and Common Voice datasets show the effectiveness and efficiency in terms of word/character error rates and real-time factors. Non-autoregressive generation and sampling via a token-wise probabilistic diffusion is demonstrated to tradeoff between accuracy and efficiency.
- 9 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 1594-1607
Dynamic Diffusion Programming and Classification
Connectionist temporal classification (CTC) has been achieving remarkable performance in utilization of neural networks to recognize sequence data like speech signal due to its simplicity, efficiency and stability. The challenging issues in CTC include assuming the conditional temporal independence and tackling the balance between accuracy and efficiency. This paper deals with these two issues by incorporating a probabilistic diffusion model into a non-autoregressive CTC for a new end-to-end automatic speech recognition (ASR). By leveraging the iterative denoising neural network for temporal classification, this method consolidates the dependency modeling in token alignments while maintaining the inference efficiency. In particular, a dynamic diffusion programming algorithm is proposed to implement an efficient computation for optimization where the diffusion objective is calculated over different alignments. Additionally, this study exploits a fast sampling strategy in reverse process, which reduces the latency by skipping those low-confidence diffusion steps. The proposed methods considerably mitigate the computation complexities in diffusion objective and sampling procedure for a discrete-diffusion ASR. The experiments on LibriSpeech and Common Voice datasets show the effectiveness and efficiency in terms of word/character error rates and real-time factors. Non-autoregressive generation and sampling via a token-wise probabilistic diffusion is demonstrated to tradeoff between accuracy and efficiency.
- 19 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 259-271
Variational Dehallucination via Retrieval-Augmented Prompt Learning
Large language models (LLMs) are powerful but prone to hallucination, generating fluent yet factually incorrect responses. Existing approaches such as prompt learning and retrieval-augmented generation (RAG) offer partial remedies in a condition that prompt learning adapts model behavior but often lacks contextual specificity while RAG improves factual grounding but treats retrieval deterministically and lacks prompt optimization. To address these gaps, we propose a variational dehallucination framework that unifies RAG with prompt learning through variational inference. By modeling prompts as latent variables, a posterior prompt encoder is learned with the condition on both questions and answers, and is aligned with a prior encoder to enable robust inference without labels. To further suppress the hallucination, a mutual information loss is incorporated to be minimized to disentangle the hidden-state features due to the prompt signals from top-$k$ and bottom $k$ retrieved examples. Experiments on Natural Questions, TriviaQA, and ASQA show that the proposed method significantly improves accuracy, semantic relevance, and consistency over strong prompting and retrieval baselines. This demonstrates the value of combining probabilistic prompt modeling with retrieval scheme to generate factual and reliable LLM outputs.
- 19 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 259-271
Variational Dehallucination via Retrieval-Augmented Prompt Learning
Large language models (LLMs) are powerful but prone to hallucination, generating fluent yet factually incorrect responses. Existing approaches such as prompt learning and retrieval-augmented generation (RAG) offer partial remedies in a condition that prompt learning adapts model behavior but often lacks contextual specificity while RAG improves factual grounding but treats retrieval deterministically and lacks prompt optimization. To address these gaps, we propose a variational dehallucination framework that unifies RAG with prompt learning through variational inference. By modeling prompts as latent variables, a posterior prompt encoder is learned with the condition on both questions and answers, and is aligned with a prior encoder to enable robust inference without labels. To further suppress the hallucination, a mutual information loss is incorporated to be minimized to disentangle the hidden-state features due to the prompt signals from top-$k$ and bottom $k$ retrieved examples. Experiments on Natural Questions, TriviaQA, and ASQA show that the proposed method significantly improves accuracy, semantic relevance, and consistency over strong prompting and retrieval baselines. This demonstrates the value of combining probabilistic prompt modeling with retrieval scheme to generate factual and reliable LLM outputs.
- 15 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33, 3896-3907
Disentangling Speech Representations Learning With Latent Diffusion for Speaker Verification
Disentangled speech representation learning for speaker verification aims to separate spoken content and speaker timbre into distinct representations. However, existing variational autoencoder (VAE)âbased methods for speech disentanglement rely on latent variables that lack semantic meaning, limiting their effectiveness for speaker verification. To address this limitation, we propose a diffusion-based method that disentangles and separates speaker features and speech content in the latent space. Building upon the VAE framework, we employ a speaker encoder to learn latent variables representing speaker features while using frame-specific latent variables to capture content. Unlike previous sequential VAE approaches, our method utilizes a conditional diffusion model in the latent space to derive speaker-aware representations. Experiments on the VoxCeleb and CN-Celeb datasets demonstrate that our method effectively isolates speaker features from speech content using pre-trained speech representations. The learned embeddings are robust to language mismatches since the speaker embeddings become content-invariant after content removal. Additionally, we design contrastive learning experiments showing that our training objective can enhance the learning of speaker-discriminative embeddings without relying on classification-based loss.
- 15 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33, 3896-3907
Disentangling Speech Representations Learning With Latent Diffusion for Speaker Verification
Disentangled speech representation learning for speaker verification aims to separate spoken content and speaker timbre into distinct representations. However, existing variational autoencoder (VAE)âbased methods for speech disentanglement rely on latent variables that lack semantic meaning, limiting their effectiveness for speaker verification. To address this limitation, we propose a diffusion-based method that disentangles and separates speaker features and speech content in the latent space. Building upon the VAE framework, we employ a speaker encoder to learn latent variables representing speaker features while using frame-specific latent variables to capture content. Unlike previous sequential VAE approaches, our method utilizes a conditional diffusion model in the latent space to derive speaker-aware representations. Experiments on the VoxCeleb and CN-Celeb datasets demonstrate that our method effectively isolates speaker features from speech content using pre-trained speech representations. The learned embeddings are robust to language mismatches since the speaker embeddings become content-invariant after content removal. Additionally, we design contrastive learning experiments showing that our training objective can enhance the learning of speaker-discriminative embeddings without relying on classification-based loss.
- Open Access
- 12 May 2025
- Published by Elsevier in Smart Agricultural Technology
- Vol. 11
Hyperspectral signature-band extraction and adaptation for sugar content prediction on Ziziphus mauritiana and Syzygium samarangense
- 12 May 2025
- Published by Elsevier in Smart Agricultural Technology
- Vol. 11
Open AccessHyperspectral signature-band extraction and adaptation for sugar content prediction on Ziziphus mauritiana and Syzygium samarangense
- Open Access
- 17 December 2024
- Published by Emerald Publishing in APSIPA Transactions on Signal and Information Processing
- Vol. 13  (1) , 1-26
Robust Multi-Domain Multi-Turn Dialogue Policy via Student-Teacher Offline Reinforcement Learning
Dialogue policy plays a crucial role in a dialogue system as it determines the system response given a user input. In a pipeline system, the dialogue policy is susceptible to the performance degradation when the preceding components fail to produce correct output. To address this issue, this paper proposes a new method to train a robust dialogue policy that can handle noisy representation due to the mispredicted user dialogue acts from natural language understanding component. This method is mainly designed with two strategies, which are student-teacher learning and offline reinforcement learning. Student-teacher learning aims to force the student model to map the extracted features of the noisy input to be close to the clean features extracted by teacher model. Meanwhile, the offline reinforcement learning with multi-label classification objective is used to train the dialogue policy to provide appropriate response given user input by only utilizing the trajectories stored in the dataset. The experimental results show that the proposed hybrid learning can substantially improve the multi-turn end-to-end performance in a pipeline dialogue using MultiWOZ 2.1 dataset under ConvLab-2 evaluation framework. Furthermore, competitive results are obtained when compared to the end-to-end performance by using the pre-trained GPT-2 model with lower computational cost.
- 17 December 2024
- Published by Emerald Publishing in APSIPA Transactions on Signal and Information Processing
- Vol. 13  (1) , 1-26
Open AccessRobust Multi-Domain Multi-Turn Dialogue Policy via Student-Teacher Offline Reinforcement Learning
Dialogue policy plays a crucial role in a dialogue system as it determines the system response given a user input. In a pipeline system, the dialogue policy is susceptible to the performance degradation when the preceding components fail to produce correct output. To address this issue, this paper proposes a new method to train a robust dialogue policy that can handle noisy representation due to the mispredicted user dialogue acts from natural language understanding component. This method is mainly designed with two strategies, which are student-teacher learning and offline reinforcement learning. Student-teacher learning aims to force the student model to map the extracted features of the noisy input to be close to the clean features extracted by teacher model. Meanwhile, the offline reinforcement learning with multi-label classification objective is used to train the dialogue policy to provide appropriate response given user input by only utilizing the trajectories stored in the dataset. The experimental results show that the proposed hybrid learning can substantially improve the multi-turn end-to-end performance in a pipeline dialogue using MultiWOZ 2.1 dataset under ConvLab-2 evaluation framework. Furthermore, competitive results are obtained when compared to the end-to-end performance by using the pre-trained GPT-2 model with lower computational cost.
- Open Access
- 4 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33  (23299290) , 111-123
Strategic Optimization for Worst-Case Augmentation and Classification
Adversarial data augmentation techniques have recently demonstrated potential in enhancing the robustness of machine learning models by identifying potential worst-case data augmentation. However, most of the existing methods implemented a single augmentation strategy for different instances through a process of greedy search, resulting in suboptimal quality of the generated data. Furthermore, previous studies that incorporated reinforcement learning (RL) to apply unique augmentation strategies required high computational cost, as it necessitated a child network to compute the reward during the optimization process. Given these limitations, this study introduces a strategic adversarial data augmentation approach that leverages RL to search for and emulate the worst-case variations through a sequence of augmentation actions. By defining a reward function with an information-theoretic perspective along with the proper definition of state space, a proficient strategy for stacking multiple augmentation strategies can be carried out in an inexpensive way and can be smoothly integrated into classifier training, thereby enhancing model robustness against unseen noises. The proposed adversarial training method was evaluated on ten different types of unseen human-readable noises across six distinct text classification tasks. Experimental results indicate that the proposed method significantly improves model robustness in compensating for unseen noises.
- 4 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 33  (23299290) , 111-123
Open AccessStrategic Optimization for Worst-Case Augmentation and Classification
Adversarial data augmentation techniques have recently demonstrated potential in enhancing the robustness of machine learning models by identifying potential worst-case data augmentation. However, most of the existing methods implemented a single augmentation strategy for different instances through a process of greedy search, resulting in suboptimal quality of the generated data. Furthermore, previous studies that incorporated reinforcement learning (RL) to apply unique augmentation strategies required high computational cost, as it necessitated a child network to compute the reward during the optimization process. Given these limitations, this study introduces a strategic adversarial data augmentation approach that leverages RL to search for and emulate the worst-case variations through a sequence of augmentation actions. By defining a reward function with an information-theoretic perspective along with the proper definition of state space, a proficient strategy for stacking multiple augmentation strategies can be carried out in an inexpensive way and can be smoothly integrated into classifier training, thereby enhancing model robustness against unseen noises. The proposed adversarial training method was evaluated on ten different types of unseen human-readable noises across six distinct text classification tasks. Experimental results indicate that the proposed method significantly improves model robustness in compensating for unseen noises.