Recent generative models have significantly advanced speech restoration tasks, yet their training objectives often misalign with human perceptual preferences, resulting in suboptimal quality. While post-training alignment has proven effective in other generative domains like text and image generation, its application to generative speech restoration remains largely under-explored. This work investigates the challenges of applying preference-based post-training to this task, focusing on how to define a robust preference signal and curate high-quality data to avoid reward hacking. To address these challenges, we propose a multi-metric preference alignment strategy. We construct a new dataset, GenSR-Pref, comprising 80K preference pairs, where each chosen sample is unanimously favored by a complementary suite of metrics covering perceptual quality, signal fidelity, content consistency, and timbre preservation. This principled approach ensures a holistic preference signal. Applying Direct Preference Optimization (DPO) with our dataset, we observe consistent and significant performance gains across three diverse generative paradigms: autoregressive models (AR), masked generative models (MGM), and flow-matching models (FM) on various restoration benchmarks, in both objective and subjective evaluations. Ablation studies confirm the superiority of our multi-metric strategy over single-metric approaches in mitigating reward hacking. Furthermore, we demonstrate that our aligned models can serve as powerful ''data annotators'', generating high-quality pseudo-labels to serve as a supervision signal for traditional discriminative models in data-scarce scenarios like singing voice restoration.

Although Speech Large Language Models have achieved notable progress, a substantial modality reasoning gap remains: their reasoning performance on speech inputs is markedly weaker than on text.This gap could be associated with representational drift across Transformer layers and behavior deviations in long-chain reasoning.To address this issue, we introduce TARS, a reinforcement-learning framework that aligns text-conditioned and speech-conditioned trajectories through an asymmetric reward design.The framework employs two dense and complementary signals: representation alignment, which measures layer-wise hidden-state similarity between speech-and text-conditioned trajectories, and behavior alignment, which evaluates semantic consistency between generated outputs and reference text completions.Experiments on challenging reasoning benchmarks, including MMSU and OBQA, show that our approach significantly narrows the modality reasoning gap and achieves state-of-the-art performance among 7B-scale Speech LLMs.

Environment sensing and fusion via onboard sensors are envisioned to be widely applied in future autonomous driving networks. This paper considers a vehicular system with multiple self-driving vehicles that is assisted by multi-access edge computing (MEC), where image data collected by the sensors is offloaded from cellular vehicles to the MEC server using vehicle-to-infrastructure (V2I) links. Sensory data can also be shared among surrounding vehicles via vehicle-to-vehicle (V2V) communication links. To improve spectrum utilization, the V2V links may reuse the same frequency spectrum as the V2I links, which may cause severe interference. To tackle this issue, we leverage reconfigurable intelligent computational surfaces (RICSs) to jointly enable V2I reflective links and mitigate interference appearing at the V2V links. Considering the limitations of traditional algorithms in addressing this problem, such as the assumption of quasi-static channel state information, which restricts their ability to adapt to dynamic environmental changes and leads to poor performance under frequently varying channel conditions, in this paper, we formulate the problem at hand as a Markov game. Our novel formulation is applied to time-varying channels subject to multi-user interference and introduces a collaborative learning mechanism among users. The considered optimization problem is solved via a driving safety-enabled multi-agent deep reinforcement learning (DS-MADRL) approach that capitalizes on the RICS presence. Our extensive numerical investigations showcase that the proposed reinforcement learning approach achieves faster convergence and significant enhancements in both data rate and driving safety, as compared to various state-of-the-art benchmarks.

Modern zero-shot text-to-speech (TTS) systems, despite using extensive pre-training, often struggle in challenging scenarios such as tongue twisters, repeated words, codeswitching, and cross-lingual synthesis, leading to intelligibility issues.To address these limitations, this paper leverages preference alignment techniques, which enable targeted construction of out-of-pretraining-distribution data to enhance performance.We introduce a new dataset, named the Intelligibility Preference Speech Dataset (INTP), and extend the Direct Preference Optimization (DPO) framework to accommodate diverse TTS architectures.After INTP alignment, in addition to intelligibility, we observe overall improvements including naturalness, similarity, and audio quality for multiple TTS models across diverse domains.Based on that, we also verify the weak-to-strong generalization ability of INTP for more intelligible models such as CosyVoice 2 and Ints.Moreover, we showcase the potential for further improvements through iterative alignment based on Ints.Audio samples are available at https://intalign.github.io/.

The recent large-scale text-to-speech (TTS) systems are usually grouped as autoregressive and non-autoregressive systems. The autoregressive systems implicitly model duration but exhibit certain deficiencies in robustness and lack of duration controllability. Non-autoregressive systems require explicit alignment information between text and speech during training and predict durations for linguistic units (e.g. phone), which may compromise their naturalness. In this paper, we introduce $\textbf{Mask}$ed $\textbf{G}$enerative $\textbf{C}$odec $\textbf{T}$ransformer (MaskGCT), a fully non-autoregressive TTS model that eliminates the need for explicit alignment information between text and speech supervision, as well as phone-level duration prediction. MaskGCT is a two-stage model: in the first stage, the model uses text to predict semantic tokens extracted from a speech self-supervised learning (SSL) model, and in the second stage, the model predicts acoustic tokens conditioned on these semantic tokens. MaskGCT follows the mask-and-predict learning paradigm. During training, MaskGCT learns to predict masked semantic or acoustic tokens based on given conditions and prompts. During inference, the model generates tokens of a specified length in a parallel manner. Experiments with 100K hours of in-the-wild speech demonstrate that MaskGCT outperforms the current state-of-the-art zero-shot TTS systems in terms of quality, similarity, and intelligibility. Audio samples are available at https://maskgct.github.io/. We release our code and model checkpoints at https://github.com/open-mmlab/Amphion/blob/main/models/tts/maskgct.

The imitation of voice, targeted on specific speech attributes such as timbre and speaking style, is crucial in speech generation. However, existing methods rely heavily on annotated data, and struggle with effectively disentangling timbre and style, leading to challenges in achieving controllable generation, especially in zero-shot scenarios. To address these issues, we propose Vevo, a versatile zero-shot voice imitation framework with controllable timbre and style. Vevo operates in two core stages: (1) Content-Style Modeling: Given either text or speech's content tokens as input, we utilize an autoregressive transformer to generate the content-style tokens, which is prompted by a style reference; (2) Acoustic Modeling: Given the content-style tokens as input, we employ a flow-matching transformer to produce acoustic representations, which is prompted by a timbre reference. To obtain the content and content-style tokens of speech, we design a fully self-supervised approach that progressively decouples the timbre, style, and linguistic content of speech. Specifically, we adopt VQ-VAE as the tokenizer for the continuous hidden features of HuBERT. We treat the vocabulary size of the VQ-VAE codebook as the information bottleneck, and adjust it carefully to obtain the disentangled speech representations. Solely self-supervised trained on 60K hours of audiobook speech data, without any fine-tuning on style-specific corpora, Vevo matches or surpasses existing methods in accent and emotion conversion tasks. Additionally, Vevo’s effectiveness in zero-shot voice conversion and text-to-speech tasks further demonstrates its strong generalization and versatility. Audio samples are available at https://versavoice.github.io/.

Though safety alignment has been applied to most large language models (LLMs), LLM service providers generally deploy a subsequent moderation as the external safety guardrail in real-world products. Existing moderators mainly practice a conventional full detection, which determines the harmfulness based on the complete LLM output, causing high service latency. Recent works pay more attention to partial detection where moderators oversee the generation midway and early stop the output if harmfulness is detected, but they directly apply moderators trained with the full detection paradigm to incomplete outputs, introducing a training-inference gap that lowers the performance. In this paper, we explore how to form a data-and-model solution that natively supports partial detection. For the data, we construct **FineHarm**, a dataset consisting of 29K prompt-response pairs with fine-grained token-level annotations to provide reasonable supervision for token-level training. Then, we propose the **Streaming Content Monitor (SCM)**, which is trained with dual supervision of response- and token-level labels and can follow the output stream of LLM to make a timely judgment of harmfulness. Experiments show that SCM gains 0.95+ in macro F1 score that is comparable to full-detection, by only seeing the first 18% of tokens in responses on average. Moreover, the SCM can serve as a pseudo-harmfulness annotator for improving safety alignment and lead to a higher harmlessness score than DPO.

Speech tokenizers serve as foundational components for speech language models, yet current designs exhibit several limitations, including: (1) dependence on multi-layer residual vector quantization structures or high frame rates, (2) reliance on auxiliary pre-trained models for semantic distillation, and (3) requirements for complex two-stage training processes. In this work, we introduce the **T**ext-**a**ware **Di**ffusion Transformer Speech **Codec** (***TaDiCodec***), a novel approach designed to overcome these challenges. TaDiCodec employs end-to-end optimization for quantization and reconstruction through a diffusion autoencoder, while integrating text guidance into the diffusion decoder to enhance reconstruction quality and achieve optimal compression. TaDiCodec achieves an extremely low frame rate of **6.25 Hz** and a corresponding bitrate of **0.0875 kbps** with a **single-layer codebook** for 24 kHz speech, while maintaining superior performance on critical speech generation evaluation metrics such as Word Error Rate (WER), speaker similarity (SIM), and speech quality (UTMOS), Notably, TaDiCodec employs a single-stage, end-to-end training paradigm, and obviating the need for auxiliary pre-trained models. We also validate the compatibility of TaDiCodec in language model based zero-shot text-to-speech with both autoregressive modeling and masked generative modeling, demonstrating its effectiveness and efficiency for speech language modeling, as well as a significantly small *reconstruction-generation gap*. To facilitate reproducibility and further research, we will make our source code and pre-trained checkpoints publicly available. Audio samples are are available at https://tadicodec.github.io/. We release code and model checkpoints at https://github.com/AmphionTeam/TaDiCodec.

We introduce ***Metis***, a foundation model for unified speech generation. Unlike previous task-specific or multi-task models, Metis follows a pre-training and fine-tuning paradigm. It is pre-trained on large-scale unlabeled speech data using masked generative modeling and then fine-tuned to adapt to diverse speech generation tasks. Specifically, (1) Metis utilizes two discrete speech representations: SSL tokens derived from speech self-supervised learning (SSL) features, and acoustic tokens directly quantized from waveforms. (2) Metis performs masked generative pre-training on SSL tokens, utilizing 300K hours of diverse speech data, without any additional condition. (3) Through fine-tuning with task-specific conditions, Metis achieves efficient adaptation to various speech generation tasks while supporting multimodal input, even when using limited data and trainable parameters. Experiments demonstrate that Metis can serve as a foundation model for unified speech generation: Metis outperforms state-of-the-art task-specific or multi-task systems across five speech generation tasks, including zero-shot text-to-speech, voice conversion, target speaker extraction, speech enhancement, and lip-to-speech, even with fewer than 20M trainable parameters or 300 times less training data. Audio samples are are available at https://metis-demo.github.io/. We release the code and model checkpoints at https://github.com/open-mmlab/Amphion.

This paper focuses on improving autonomous driving safety via task offloading from cellular vehicles (CVs), using vehicle-to-infrastructure (V2I) links, to a multi-access edge computing (MEC) server. Considering that the V2I links sometimes can be reused by vehicle-to-vehicle (V2V) communications to improve spectrum utilization, the receiver of the V2I link may suffer from severe interference, causing outages during the task offloading. To tackle this issue, we propose the deployment of a reconfigurable intelligent computational surface (RICS) to enable, not only V2I reflective links but also interference cancellation at the V2V links exploiting the computational capability of its metamaterials. We devise a joint optimization formulation for the task offloading ratio between the CVs and the MEC server, the spectrum sharing strategy between V2V and V2I communications, as well as the RICS reflection and refraction matrices, to maximize a safety-based autonomous driving task. Due to the non-convexity of the problem and the coupling among its free variables, we transform it into a more tractable equivalent form, which is then decomposed into three sub-problems and solved via an alternate approximation method. Simulation results show that the proposed RICS-assisted offloading framework significantly improves the safety of the autonomous driving network, in which the safety coefficient of the CVs is improved by nearly 34%. The V2V data rate is improved by around 60%, which indicates that the RICS’s adjustment of the signals can effectively mitigate the interference of the V2V link.

Conversational humor is the key to capturing dialogue semantics and dialogue comprehension, which is usually generated in multiple modalities, such as linguistic rhetoric (textual modality), exaggerated facial expressions or movements (visual modality), and quirky intonation (acoustic modality). However, existing multimodal corpora for conversation humor are coarse-grained, and the modality is insufficient to support the conversational humor recognition task. This paper designed an annotation scheme for multimodal humor datasets, and constructed a corpus based on a Chinese sitcom for conversational humor recognition, named MUCH. The MUCH corpus consists of 34,804 utterances in total, and 7,079 of them are humorous. We employed both unimodal and multimodal methods to test our MUCH corpus. Experimental results showed that the multimodal approach could achieve 75.94% in terms of F1-score and surpassed the performance of most unimodal methods, which demonstrated that the MUCH corpus was effective for multimodal humor recognition tasks.

Generative Adversarial Network (GAN) based vocoders are superior in inference speed and synthesis quality when reconstructing an audible waveform from an acoustic representation. This study focuses on improving the discriminator to promote GAN-based vocoders. Most existing time-frequency-representation-based discriminators are rooted in Short-Time Fourier Transform (STFT), whose time-frequency resolution in a spectrogram is fixed, making it incompatible with signals like singing voices that require flexible attention for different frequency bands. Motivated by that, our study utilizes the Constant-Q Transform (CQT), which owns dynamic resolution among frequencies, contributing to a better modeling ability in pitch accuracy and harmonic tracking. Specifically, we propose a Multi-Scale Sub-Band CQT (MS-SB-CQT) Discriminator, which operates on the CQT spectrogram at multiple scales and performs sub-band processing according to different octaves. Experiments conducted on both speech and singing voices confirm the effectiveness of our proposed method. Moreover, we also verified that the CQT-based and the STFT-based discriminators could be complementary under joint training. Specifically, enhanced by the proposed MS-SB-CQT and the existing MS-STFT Discriminators, the MOS of HiFi-GAN can be boosted from 3.27 to 3.87 for seen singers and from 3.40 to 3.78 for unseen singers.

Singing Voice Conversion (SVC) is a technique that enables any singer to perform any song. To achieve this, it is essential to obtain speaker-agnostic representations from the source audio, which poses a significant challenge. A common solution involves utilizing a semantic-based audio pretrained model as a feature extractor However, the degree to which the extracted features can meet the SVC requirements remains an open question. This includes their capability to accurately model melody and lyrics, the speaker-independency of their underlying acoustic information, and their robustness for in-the-wild acoustic environments. In this study, we investigate the knowledge within classical semantic-based pretrained models in much detail. We discover that the knowledge of different models is diverse and can be complementary for SVC. Based on the above, we design a Singing Voice Conversion framework based on Diverse Semantic-based Feature Fusion (DSFF-SVC). Experimental results demonstrate that DSFF-SVC can be generalized and improve various existing SVC models, particularly in challenging real-world conversion tasks. Our demo website is available at https://diversesemanticsvc.github.io/.

Amphion is an open-source toolkit for Audio, Music, and Speech Generation, targeting to ease the way for junior researchers and engineers into these fields. It presents a unified framework that includes diverse generation tasks and models, with the added bonus of being easily extendable for new incorporation. The toolkit is designed with beginner-friendly workflows and pre-trained models, allowing both beginners and seasoned researchers to kick-start their projects with relative ease. The initial release of Amphion v0.1 supports a range of tasks including Text to Speech (TTS), Text to Audio (TTA), and Singing Voice Conversion (SVC), supplemented by essential components like data preprocessing, state-of-the-art vocoders, and evaluation metrics. This paper presents a high-level overview of Amphion. Amphion is open-sourced at https://github.com/open-mmlab/Amphion.

In this paper, we focus on improving autonomous driving safety via task offloading from cellular vehicles (CVs), using vehicle-to-infrastructure (V2I) links, to an multi-access edge computing (MEC) server. Considering that the frequencies used for V2I links can be reused for vehicle-to-vehicle (V2V) communications to improve spectrum utilization, the receiver of each V2I link may suffer from severe interference, causing outages in the task offloading process. To tackle this issue, we propose the deployment of a reconfigurable intelligent computational surface (RICS) to enable, not only V2I reflective links, but also interference cancellation at the V2V links exploiting the computational capability of its metamaterials. We devise a joint optimization formulation for the task offloading ratio between the CVs and the MEC server, the spectrum sharing strategy between V2V and V2I communications, as well as the RICS reflection and refraction matrices, with the objective to maximize a safety-based autonomous driving task. Due to the non-convexity of the problem and the coupling among its free variables, we transform it into a more tractable equivalent form, which is then decomposed into three sub-problems and solved via an alternate approximation method. Our simulation results demonstrate the effectiveness of the proposed RICS optimization in improving the safety in autonomous driving networks.

Generative Adversarial Network (GAN) based vocoders are superior in both inference speed and synthesis quality when reconstructing an audible waveform from an acoustic representation. This study focuses on improving the discriminator for GAN-based vocoders. Most existing Time-Frequency Representation (TFR)-based discriminators are rooted in Short-Time Fourier Transform (STFT), which owns a constant Time-Frequency (TF) resolution, linearly scaled center frequencies, and a fixed decomposition basis, making it incompatible with signals like singing voices that require dynamic attention for different frequency bands and different time intervals. Motivated by that, we propose a Multi-Scale Sub-Band Constant-Q Transform CQT (MS-SB-CQT) discriminator and a Multi-Scale Temporal-Compressed Continuous Wavelet Transform CWT (MS-TC-CWT) discriminator. Both CQT and CWT have a dynamic TF resolution for different frequency bands. In contrast, CQT has a better modeling ability in pitch information, and CWT has a better modeling ability in short-time transients. Experiments conducted on both speech and singing voices confirm the effectiveness of our proposed discriminators. Moreover, the STFT, CQT, and CWT-based discriminators can be used jointly for better performance. The proposed discriminators can boost the synthesis quality of various state-of-the-art GAN-based vocoders, including HiFi-GAN, BigVGAN, and APNet.

With the continuous development of mobile communication networks and artificial intelligence technology, the number of smart mobile devices has shown an exponential growth trend, and artificial intelligence (AI) jobs have developed unprecedentedly. However, it is difficult for resource-constrained mobile devices to meet the computational demands of these jobs. How to make full use of the dynamic resources in the wireless network to achieve efficient execution of AI jobs is the evolution direction of the next-generation network. To achieve this goal, we model the job as a directed acyclic graph (DAG), partition it into executors based on the type of task, and minimize the execution time of all jobs in 6G wireless networks by optimizing executors deployment. Considering the dynamic features of channel states and DAG topology, the optimization problem is addressed by deep reinforcement learning, i.e., Deep Q-Network (DQN). In the simulation, we manifest the performance of the DQN-based DAG task scheduling in terms of convergence and latency.

Fake news detection is crucial for preventing the dissemination of misinformation on social media.To differentiate fake news from real ones, existing methods observe the language patterns of the news post and "zoom in" to verify its content with knowledge sources or check its readers' replies.However, these methods neglect the information in the external news environment where a fake news post is created and disseminated.The news environment represents recent mainstream media opinion and public attention, which is an important inspiration of fake news fabrication because fake news is often designed to ride the wave of popular events and catch public attention with unexpected novel content for greater exposure and spread.To capture the environmental signals of news posts, we "zoom out" to observe the news environment and propose the News Environment Perception Framework (NEP).For each post, we construct its macro and micro news environment from recent mainstream news.Then we design a popularity-oriented and a noveltyoriented module to perceive useful signals and further assist final prediction.Experiments on our newly built datasets show that the NEP can efficiently improve the performance of basic fake news detectors.1

Pop music generation has always been an attractive topic for both musicians and scientists for a long time. However, automatically composing pop music with a satisfactory structure is still a challenging issue. In this paper, we propose to leverage harmony-aware learning for structure-enhanced pop music generation. On the one hand, one of the participants of harmony, chord, represents the harmonic set of multiple notes, which is integrated closely with the spatial structure of music, the texture. On the other hand, the other participant of harmony, chord progression, usually accompanies the development of the music, which promotes the temporal structure of music, the form. Moreover, when chords evolve into chord progression, the texture and form can be bridged by the harmony naturally, which contributes to the joint learning of the two structures. Furthermore, we propose the Harmony-Aware Hierarchical Music Transformer (HAT), which can exploit the structure adaptively from the music, and make the musical tokens interact hierarchically to enhance the structure in multi-level musical elements. Experimental results reveal that compared to the existing methods, HAT owns a much better understanding of the structure and it can also improve the quality of generated music, especially in the form and texture.

Qiang Sheng, Juan Cao, Xueyao Zhang, Xirong Li, Lei Zhong. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

To defend against fake news, researchers have developed various methods based on texts. These methods can be grouped as 1) pattern-based methods, which focus on shared patterns among fake news posts rather than the claim itself; and 2) fact-based methods, which retrieve from external sources to verify the claim's veracity without considering patterns. The two groups of methods, which have different preferences of textual clues, actually play complementary roles in detecting fake news. However, few works consider their integration. In this paper, we study the problem of integrating pattern- and fact-based models into one framework via modeling their preference differences, i.e., making the pattern- and fact-based models focus on respective preferred parts in a post and mitigate interference from non-preferred parts as possible. To this end, we build a Preference-aware Fake News Detection Framework (Pref-FEND), which learns the respective preferences of pattern- and fact-based models for joint detection. We first design a heterogeneous dynamic graph convolutional network to generate the respective preference maps, and then use these maps to guide the joint learning of pattern- and fact-based models for final prediction. Experiments on two real-world datasets show that Pref-FEND effectively captures model preferences and improves the performance of models based on patterns, facts, or both.

Emotion plays an important role in detecting fake news online. When leveraging emotional signals, the existing methods focus on exploiting the emotions of news contents that conveyed by the publishers (i.e., publisher emotion). However, fake news often evokes high-arousal or activating emotions of people, so the emotions of news comments aroused in the crowd (i.e., social emotion) should not be ignored. Furthermore, it remains to be explored whether there exists a relationship between publisher emotion and social emotion (i.e., dual emotion), and how the dual emotion appears in fake news. In this paper, we verify that dual emotion is distinctive between fake and real news and propose Dual Emotion Features to represent dual emotion and the relationship between them for fake news detection. Further, we exhibit that our proposed features can be easily plugged into existing fake news detectors as an enhancement. Extensive experiments on three real-world datasets (one in English and the others in Chinese) show that our proposed feature set: 1) outperforms the state-of-the-art task-related emotional features; 2) can be well compatible with existing fake news detectors and effectively improve the performance of detecting fake news.1 2