Wenxi Chen, Ruiqi Yan, Yushen Chen, Zhikang Niu, Ziyang Ma, Xiquan Li, Yuzhe Liang, Wenhanlin, Shunshun Yin, Ming Tao, Xinsheng Wang, Xie Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Chunyu Qiang, Xiaopeng Wang, Kang Yin, Yuzhe Liang, Yuxin Guo, Teng Ma, Ziyu Zhang, Tianrui Wang, Cheng Gong, Yushen Chen, Ruibo Fu, Longbiao Wang, Jianwu Dang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

This paper introduces F5-TTS, a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT). Without requiring complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler tokens to the same length as input speech, and then the denoising is performed for speech generation, which was originally proved feasible by E2 TTS. However, the original design of E2 TTS makes it hard to follow due to its slow convergence and low robustness. To address these issues, we first model the input with ConvNeXt to refine the text representation, making it easy to align with the speech. We further propose an inference-time Sway Sampling strategy, which significantly improves our model’s performance and efficiency. This sampling strategy for flow step can be easily applied to existing flow matching based models without retraining. Our design allows faster training and achieves an inference RTF of 0.15, which is greatly improved compared to state-of-the-art diffusion-based TTS models. Trained on a public 100K hours multilingual dataset, our F5-TTS exhibits highly natural and expressive zero-shot ability, seamless code-switching capability, and speed control efficiency. We have released all codes and checkpoints to promote community development, at https://SWivid.github.io/F5-TTS/.

Federated distillation (FD), which inherits the privacy-preserving nature of federated learning (FL), has recently attracted increasing attention due to its communication efficiency in training the global model through logits aggregation. However, FD performance suffers from slow convergence due to statistically heterogeneous data and failures in logits transmission from resource-constrained clients over unreliable wireless links. Client selection and efficient resource allocation are typical methods to speed up the convergence in FD. However, most existing works have not considered the inherent inter-dependencies between these two methods. To address this issue, in this paper, we propose a Stackelberg game-based framework to optimally balance client selection and resource allocation. Specifically, client selection is formulated as the leader-level problem to reduce the number of required communication rounds. Subsequently, resource allocation is formulated as the follower-level problem to maximize their successful uploading rates in each round. By decomposing the follower-level problem into three subproblems, the closed-form solutions of transmission power, computation frequency, and the number of uploaded logits allocations are derived through monotonicity analysis. To solve the leader-level problem, we first derive the upper bound of the convergence of the FD global loss. Based on this, an uncertainty-based client selection scheme and an attention-based logits sampling method are proposed to optimally solve the leader’s optimization problem. Finally, the Stackelberg equilibrium is reached when all selected clients can successfully upload logits to the server. Simulation results demonstrate that the proposed Stackelberg equilibrium solutions significantly enhance the global model convergence speed.

The exponential growth of the Internet of Things (IoT) ecosystems has raised significant cybersecurity concerns. Deep learning (DL)-based methods have shown promising performance in detecting potential cyber threats in IoT networks. However, as these methods often involve data centralization, they can pose serious data privacy issues for IoT users and increase the communication burden of local networks. Federated learning (FL), as a distributed learning paradigm, enables privacy-preserving training of IoT intrusion detection models by requiring only model updates from IoT devices. However, the resource-constrained nature of IoT devices can significantly decrease FL training efficiencies, such as increased training latency and delayed convergence speed. Moreover, the data heterogeneous issues of IoT devices can also impact the accuracy and robustness of the trained model. To address these challenges, we propose an efficient FL framework, FedKD-Prox, based on federated proximal (FedProx) and knowledge distillation (KD). To improve the prediction accuracy within a limited time budget, the proposed framework aims to efficiently exploit the computation capability of the IoT trainers, reduce the communication overhead of FL, and alleviate the impact of heterogeneous data issues. The simulation results show that FedKD-Prox achieves higher accuracy and improves the robustness of the trained intrusion detection model.