论文检索

Cunhang Fan, Enrui Liu, Andong Li, Jianhua Tao, Jian Zhou, Jiahao Li, Chengshi Zheng, Zhao Lv

Although the complex spectrum-based speech enhancement (SE) methods have achieved significant performance, coupling amplitude and phase can lead to a compensation effect, where amplitude information is sacrificed to compensate for the phase that is harmful to SE. In addition, to further improve the performance of SE, many modules are stacked onto SE, resulting in increased model complexity that limits the application of SE. To address these problems, we proposed a dual-path network based on compressed frequency using Mamba. First, we extract amplitude and phase information through parallel dual branches. This approach leverages structured complex spectra to implicitly capture phase information and solves the compensation effect by decoupling amplitude and phase, and the network incorporates an interaction module to suppress unnecessary parts and recover missing components from the other branch. Second, to reduce network complexity, the network introduces a band-split strategy to compress the frequency dimension. To further reduce complexity while maintaining good performance, we designed a Mamba-based module that models the time and frequency dimensions under linear complexity. Finally, compared to baselines, our model achieves an average 8.3 times reduction in computational complexity while maintaining superior performance. Furthermore, it achieves a 25 times reduction in complexity compared to transformer-based models.

Jianwei Cui, Yu Gu, Shihao Chen, Jie Zhang, Liping Chen, Lirong Dai

Singing Voice Synthesis (SVS) aims to generate singing voices of high fidelity and expressiveness. Conventional SVS systems usually utilize an acoustic model to transform a music score into acoustic features, followed by a vocoder to reconstruct the singing voice. It was recently shown that end-to-end modeling is effective in the fields of SVS and Text to Speech (TTS). In this work, we thus present a fully end-to-end SVS method together with a chunkwise streaming inference to address the latency issue for practical usages. Note that this is the first attempt to fully implement end-to-end streaming audio synthesis using latent representations in VAE. We have made specific improvements to enhance the performance of streaming SVS using latent representations. Experimental results demonstrate that the proposed method achieves synthesized audio with high expressiveness and pitch accuracy in both streaming SVS and TTS tasks.

Yujie Chen, Jiangyan Yi, Cunhang Fan, Jianhua Tao, Yong Ren, Siding Zeng, Chu Yuan Zhang, Xinrui Yan, Hao Gu, Jun Xue 等

Rapid advancements in speech synthesis and voice conversion bring convenience but also new security risks, creating an urgent need for effective audio deepfake detection. Although current models perform well, their effectiveness diminishes when confronted with the diverse and evolving nature of real-world deepfakes. To address this issue, we propose a continual learning method named Region-Based Optimization (RegO) for audio deepfake detection. Specifically, we use the Fisher information matrix to measure important neuron regions for real and fake audio detection, dividing them into four regions. First, we directly fine-tune the less important regions to quickly adapt to new tasks. Next, we apply gradient optimization in parallel for regions important only to real audio detection, and in orthogonal directions for regions important only to fake audio detection. For regions that are important to both, we use sample proportion-based adaptive gradient optimization. This region-adaptive optimization ensures an appropriate trade-off between memory stability and learning plasticity. Additionally, to address the increase of redundant neurons from old tasks, we further introduce the Ebbinghaus forgetting mechanism to release them, thereby promoting the model’s ability to learn more generalized discriminative features. Experimental results show our method achieves a 21.3 percent improvement in EER over the state-of-the-art continual learning approach RWM for audio deepfake detection. Moreover, the effectiveness of RegO extends beyond the audio deepfake detection domain, showing potential significance in other tasks, such as image recognition.

Li Chai, Donglin Wang

Lyric-to-melody generation is a highly challenging task in the field of AI music generation. Due to the difficulty of learning strict yet weak correlations between lyrics and melodies, previous methods have suffered from weak controllability, low-quality and poorly structured generation. To address these challenges, we propose CSL-L2M, a controllable song-level lyric-to-melody generation method based on an in-attention Transformer decoder with fine-grained lyric and musical controls, which is able to generate full-song melodies matched with the given lyrics and user-specified musical attributes. Specifically, we first introduce REMI-Aligned, a novel music representation that incorporates strict syllable- and sentence-level alignments between lyrics and melodies, facilitating precise alignment modeling. Subsequently, sentence-level semantic lyric embeddings independently extracted from a sentence-wise Transformer encoder are combined with word-level part-of-speech embeddings and syllable-level tone embeddings as fine-grained controls to enhance the controllability of lyrics over melody generation. Then we introduce human-labeled musical tags, sentence-level statistical musical attributes, and learned musical features extracted from a pre-trained VQ-VAE as coarse-grained, fine-grained and high-fidelity controls, respectively, to the generation process, thereby enabling user control over melody generation. Finally, an in-attention Transformer decoder technique is leveraged to exert fine-grained control over the full-song melody generation with the aforementioned lyric and musical conditions. Experimental results demonstrate that our proposed CSL-L2M outperforms the state-of-the-art models, generating melodies with higher quality, better controllability and enhanced structure.

Kuluhan Binici, Abhinav Ramesh Kashyap, Viktor Schlegel, Andy T. Liu, Vijay Prakash Dwivedi, Thanh-Tung Nguyen, Xiaoxue Gao, Nancy F. Chen, Stefan Winkler

Automatic Speech Recognition (ASR) systems are pivotal in transcribing speech into text, yet the errors they introduce can significantly degrade the performance of downstream tasks like summarization. This issue is particularly pronounced in clinical dialogue summarization, a low-resource domain where supervised data for fine-tuning is scarce, necessitating the use of ASR models as black-box solutions. Employing conventional data augmentation for enhancing the noise robustness of summarization models is not feasible either due to the unavailability of sufficient medical dialogue audio recordings and corresponding ASR transcripts. To address this challenge, we propose MEDSAGE, an approach for generating synthetic samples for data augmentation using Large Language Models (LLMs). Specifically, we leverage the in-context learning capabilities of LLMs and instruct them to generate ASR-like errors based on a few available medical dialogue examples with audio recordings. Experimental results show that LLMs can effectively model ASR noise, and incorporating this noisy data into the training process significantly improves the robustness and accuracy of medical dialogue summarization systems. This approach addresses the challenges of noisy ASR outputs in critical applications, offering a robust solution to enhance the reliability of clinical dialogue summarization.

Zhe Ye, Wenjie Zhang, Ying Ren, Xiangui Kang, Diqun Yan, Bin Ma, Shiqi Wang

Backdoor attacks pose a significant threat during the model's training phase. Attackers craft pre-defined triggers to break deep neural networks, ensuring the model accurately classifies clean samples during inference yet erroneously classifies samples added with these triggers. Recent studies have shown that speaker recognition systems trained on large-scale data are susceptible to backdoor attacks. Existing attackers employ unnoticed ambient sounds as triggers. However, these sounds are not inherently part of the training samples themselves. In essence, triggers can be designed to maintain an intrinsic connection with the original speech to enhance stealthiness. Our paper presents a novel attack methodology named Speed Master, which undermines deep neural networks by manipulating the speed of speech samples. Specifically, we execute poison-only backdoor attacks using speed or tempo adjustment. Changes in speech rate have become a common occurrence, as seen on platforms that allow users to adjust playback speed. In real-world scenarios, people naturally adjust their speaking rate depending on the context. As a result, changes in a speaker’s speech rate are typically perceived as normal and are unlikely to raise suspicion. Furthermore, detecting such subtle adjustments becomes challenging for users without reference speech. Our comprehensive experiments demonstrate that Speed Master can achieve an ASR over 99% in the digital domain, with only a 0.6% poisoning rate. Additionally, we validate the feasibility of Speed Master in the real world and its resistance to typical defensive measures.

Bartosz Wójcik, Alessio Devoto, Karol Pustelnik, Pasquale Minervini, Simone Scardapane

While transformer models have been highly successful, they are computationally inefficient. We observe that for each layer, the full width of the layer may be needed only for a small subset of tokens inside a batch and that the "effective" width needed to process a token can vary from layer to layer. Motivated by this observation, we introduce the Adaptive Computation Module (ACM), a generic module that dynamically adapts its computational load to match the estimated difficulty of the input on a per-token basis. An ACM consists of a sequence of learners that progressively refine the output of their preceding counterparts. An additional gating mechanism determines the optimal number of learners to execute for each token. We also propose a distillation technique to replace any pre-trained model with an "ACMized" variant. Our evaluation of transformer models in computer vision and speech recognition demonstrates that substituting layers with ACMs significantly reduces inference costs without degrading the downstream accuracy for a wide interval of user-defined budgets.

Yi Li, Yang Sun, Plamen P Angelov

In this paper, we present a novel diffusion model-based monaural speech enhancement method. Our approach incorporates the separate estimation of speech spectra's magnitude and phase in two diffusion networks. Throughout the diffusion process, noise clips from real-world noise interferences are added gradually to the clean speech spectra and a noise-aware reverse process is proposed to learn how to generate both clean speech spectra and noise spectra. Furthermore, to fully leverage the intrinsic relationship between magnitude and phase, we introduce a complex-cycle-consistent (CCC) mechanism that uses the estimated magnitude to map the phase, and vice versa. We implement this algorithm within a phase-aware speech enhancement diffusion model (SEDM). We conduct extensive experiments on public datasets to demonstrate the effectiveness of our method, highlighting the significant benefits of exploiting the intrinsic relationship between phase and magnitude information to enhance speech. The comparison to conventional diffusion models demonstrates the superiority of SEDM.

Xuanming Jiang, Baoyi An, Guoshuai Zhao, Xueming Qian

Audio classification plays a crucial role within fields such as human-machine interaction and intelligent robotics. However, high-performance audio classification systems typically demand significant computational and storage resources, posing substantial challenges when deploying to the resource-constrained edge devices with an urgent need for such capabilities. To achieve a new level of balance between model complexity and performance, we introduce a novel multi-view method for the separated time-frequency features extraction and utilization, which exists within the proposed Mini Mirror Multi-View Network (M3Net) in the form of the Mirror Attention mechanism. M3Net enables reversible spatial transformation of spectral features is capable of efficiently leverages robust local and global features in the time and frequency domains with low requirements for parameters. Experiments based on Mel-Spectrogram without data augmentation and pre-training indicate that M3Net can achieve classification accuracy over 97% on the UrbanSound8K and SpeechCommandsV2 datasets with only 0.03 million parameters. The contribution of each functional segment in M3Net is fully verified and explained in the ablation experiments.

Sujan Dutta, Deepak Pandita, Tharindu Cyril Weerasooriya, Marcos Zampieri, Christopher M. Homan, Ashiqur R. KhudaBukhsh

Ensuring annotator quality in training and evaluation data is a key piece of machine learning in NLP. Tasks such as sentiment analysis and offensive speech detection are intrinsically subjective, creating a challenging scenario for traditional quality assessment approaches because it is hard to distinguish disagreement due to poor work from that due to differences of opinions between sincere annotators. With the goal of increasing diverse perspectives in annotation while ensuring consistency, we propose ARTICLE, an in-context learning (ICL) framework to estimate annotation quality through self-consistency. We evaluate this framework on two offensive speech datasets using multiple LLMs and compare its performance with traditional methods. Our findings indicate that ARTICLE can be used as a robust method for identifying reliable annotators, hence improving data quality.

Yifan Xie, Tao Feng, Xin Zhang, Xiangyang Luo, Zixuan Guo, Weijiang Yu, Heng Chang, Fei Ma, Fei Richard Yu

Talking head synthesis with arbitrary speech audio is a crucial challenge in the field of digital humans. Recently, methods based on radiance fields have received increasing attention due to their ability to synthesize high-fidelity and identity-consistent talking heads from just a few minutes of training video. However, due to the limited scale of the training data, these methods often exhibit poor performance in audio-lip synchronization and visual quality. In this paper, we propose a novel 3D Gaussian-based method called PointTalk, which constructs a static 3D Gaussian field of the head and deforms it in sync with the audio. It also incorporates an audio-driven dynamic lip point cloud as a critical component of the conditional information, thereby facilitating the effective synthesis of talking heads. Specifically, the initial step involves generating the corresponding lip point cloud from the audio signal and capturing its topological structure. The design of the dynamic difference encoder aims to capture the subtle nuances inherent in dynamic lip movements more effectively. Furthermore, we integrate the audio-point enhancement module, which not only ensures the synchronization of the audio signal with the corresponding lip point cloud within the feature space, but also facilitates a deeper understanding of the interrelations among cross-modal conditional features. Extensive experiments demonstrate that our method achieves superior high-fidelity and audio-lip synchronization in talking head synthesis compared to previous methods.

Youjia Wang, Yiwen Wu, Hengan Zhou, Hongyang Lin, Xingyue Peng, Jingyan Zhang, Yingsheng Zhu, YingWenQi Jiang, Yatu Zhang, Lan Xu 等

We present Capturing the Unseen (CAPUS), a novel facial motion capture (MoCap) technique that operates without visual signals. CAPUS leverages miniaturized Inertial Measurement Units (IMUs) as a new sensing modality for facial motion capture. While IMUs have become essential in full-body MoCap for their portability and independence from environmental conditions, their application in facial MoCap remains underexplored. We address this by customizing micro-IMUs, small enough to be placed on the face, and strategically positioning them in alignment with key facial muscles to capture expression dynamics. CAPUS introduces the first facial IMU dataset, encompassing both IMU and visual signals from participants engaged in diverse activities such as multilingual speech, facial expressions, and emotionally intoned auditions. We train a Transformer Diffusion-based neural network to infer Blendshape parameters directly from IMU data. Our experimental results demonstrate that CAPUS reliably captures facial motion in conditions where visual-based methods struggle, including facial occlusions, rapid movements, and low-light environments. Additionally, by eliminating the need for visual inputs, CAPUS offers enhanced privacy protection, making it a robust solution for various applications.

Tao Liu, Ziyang Ma, Qi Chen, Feilong Chen, Shuai Fan, Xie Chen, Kai Yu

We present VQTalker, a Vector Quantization-based framework for multilingual talking head generation that addresses the challenges of lip synchronization and natural motion across diverse languages. Our approach is grounded in the phonetic principle that human speech comprises a finite set of distinct sound units (phonemes) and corresponding visual articulations (visemes), which often share commonalities across languages. We introduce a facial motion tokenizer based on Group Residual Finite Scalar Quantization (GRFSQ), which creates a discretized representation of facial features. This method enables comprehensive capture of facial movements while improving generalization to multiple languages, even with limited training data. Building on this quantized representation, we implement a coarse-to-fine motion generation process that progressively refines facial animations. Extensive experiments demonstrate that VQTalker achieves state-of-the-art performance in both video-driven and speech-driven scenarios, particularly in multilingual settings. Notably, our method achieves high-quality results at a resolution of 512 × 512 pixels while maintaining a lower bitrate of approximately 11 kbps. Our work opens new possibilities for cross-lingual talking face generation.

Zonglin Li, Xiaoqian Lv, Qinglin Liu, Quanling Meng, Xin Sun, Shengping Zhang

Most existing 3D visual speech animation methods synthesize lip movements synchronized with speech, which however neglect head poses and therefore degrade the animation realism. The animation of head poses presents two primary challenges: (1) the intricate mapping between speech and head poses remains poorly understood and (2) the absence of 4D face datasets featuring realistic head poses. Inspired by prosody decomposition in speech processing, we discern that head movements correlate with the fundamental frequency (F0) of speech prosody, while lip movements align with the language content. These observations motivate us to propose a novel framework, dubbed ProsodyTalker, that concurrently synthesizes lip and head movements, grounded in the principles of prosody decomposition. The core idea is first to adopt information perturbation to explicitly decompose the speech prosody into pose-related F0 and lip-related language content. Then, an autoregressive content-oriented fusion decoder is employed to enhance lip synchronization in the synthesized facial sequences. To synthesize head poses, we design a transformer-based variational autoencoder to learn a latent distribution of facial sequences and propose an F0-conditioned latent diffusion model to establish a probabilistic mapping from F0 to pose-related latent codes. Furthermore, we contribute a large-scale 4D face dataset containing bunches of variations in identities, head poses and facial motions. Extensive experiments show that our method achieves more realistic animation than state-of-the-art methods.

Jisoo Kim, Jungbin Cho, Joonho Park, Soonmin Hwang, Da Eun Kim, Geon Kim, Youngjae Yu

Speech-driven 3D facial animation has garnered lots of attention thanks to its broad range of applications. Despite recent advancements in achieving realistic lip motion, current methods fail to capture the nuanced emotional undertones conveyed through speech and produce monotonous facial motion. These limitations result in blunt and repetitive facial animations, reducing user engagement and hindering their applicability. To address these challenges, we introduce DEEPTalk, a novel approach that generates diverse and emotionally rich 3D facial expressions directly from speech inputs. To achieve this, we first train DEE (Dynamic Emotion Embedding), which employs probabilistic contrastive learning to forge a joint emotion embedding space for both speech and facial motion. This probabilistic framework captures the uncertainty in interpreting emotions from speech and facial motion, enabling the derivation of emotion vectors from its multifaceted space. Moreover, to generate dynamic facial motion, we design TH-VQVAE (Temporally Hierarchical VQ-VAE) as an expressive and robust motion prior overcoming limitations of VAEs and VQ-VAEs. Utilizing these strong priors, we develop DEEPTalk, a talking head generator that non-autoregressively predicts codebook indices to create dynamic facial motion, incorporating a novel emotion consistency loss. Extensive experiments on various datasets demonstrate the effectiveness of our approach in creating diverse, emotionally expressive talking faces that maintain accurate lip-sync. Our project page is available at https://whwjdqls.github.io/deeptalk.github.io/.

Jun Gao, Qian Qiao, Tianxiang Wu, Zili Wang, Ziqiang Cao, Wenjie Li

In-context learning (ICL) advances Large Language Models (LLMs) exhibiting emergent ability on downstream tasks without updating billions of parameters. However, in the area of multimodal Large Language Models (MLLMs), two problems hinder the application of multimodal ICL: (1) Most primary MLLMs are only trained on single-image datasets, making them unable to read extra multimodal demonstrations. (2) With the demonstrations increasing, thousands of visual tokens highly challenge hardware and degrade ICL performance. During preliminary explorations, we discovered that the inner LLM focuses more on the linguistic modality within multimodal demonstrations during generation. Therefore, we propose a general and lightweight framework AIM to tackle the mentioned problems through Aggregating Image information of Multimodal demonstrations to the latent space of the corresponding textual labels. After aggregation, AIM substitutes each demonstration with generated fused virtual tokens whose length is reduced to the same as its texts. Except for shortening input length, AIM further upgrades MLLMs pre-trained on image-text pairs to support multimodal ICL, as images from demonstrations are disregarded. Furthermore, benefiting from aggregating different demonstrations independently, AIM configures Demonstration Bank (DB) to avoid repeated aggregation, which significantly boosts model efficiency. We build AIM upon QWen-VL and LLaVA-Next, and AIM is comprehensively evaluated on image caption, VQA, and hateful speech detection. Outstanding results reveal that AIM provides an efficient and effective solution in upgrading MLLMs for multimodal ICL.

Yongkang Cheng, Shaoli Huang, Xuelin Chen, Jifeng Ning, Mingming Gong

Diffusion models have demonstrated remarkable synthesis quality and diversity in generating co-speech gestures. However, the computationally intensive sampling steps associated with diffusion models hinder their practicality in real-world applications. Hence, we present DIDiffGes, for a Decoupled Semi-Implicit Diffusion model-based framework, that can synthesize high-quality, expressive gestures from speech using only a few sampling steps. Our approach leverages Generative Adversarial Networks (GANs) to enable large-step sampling for diffusion model. We decouple gesture data into body and hands distributions and further decompose them into marginal and conditional distributions. GANs model the marginal distribution implicitly, while L2 reconstruction loss learns the conditional distributions exciplictly. This strategy enhances GAN training stability and ensures expressiveness of generated full-body gestures. Our framework also learns to denoise root noise conditioned on local body representation, guaranteeing stability and realism. DIDiffGes can generate gestures from speech with just 10 sampling steps, without compromising quality and expressiveness, reducing the number of sampling steps by a factor of 100 compared to existing methods. Our user study reveals that our method outperforms state-of-the-art approaches in human likeness, appropriateness, and style correctness.

Yuxuan Bian, Ailing Zeng, Xuan Ju, Xian Liu, Zhaoyang Zhang, Wei Liu, Qiang Xu

Whole-body multimodal motion generation, controlled by text, speech, or music, has numerous applications including video generation and character animation. However, employing a unified model to process different condition modalities presents two main challenges: motion distribution drifts across different tasks (e.g., co-speech gestures and text-driven daily actions) and the complex optimization of mixed conditions with varying granularities (e.g., text and audio). In this paper, we propose MotionCraft, a unified diffusion transformer that crafts whole-body motion with plug-and-play multimodal control. Our framework employs a coarse-to-fine training strategy, starting with the text-to-motion semantic pre-training, followed by the multimodal low-level control adaptation. To effectively learn and transfer motion knowledge across different distributions, we design MC-Attn for parallel modeling of static and dynamic human topology graphs. To overcome the motion format inconsistency of existing benchmarks, we introduce MC-Bench, the first available multimodal whole-body motion generation benchmark based on the unified SMPL-X format. Extensive experiments show that MotionCraft achieves state-of-the-art performance on various standard motion generation tasks.

Yuxuan Song, Qiudan Li, Yilin Wu, David Jingjun Xu, Daniel Dajun Zeng

Personality identification plays important roles in understanding user behavior and offering foresight ability for downstream applications. The key challenge is how to address the scarcity of labeled personality data. Recently, some studies have adopted data augmentation and prompt learning to perform personality identification. However, they still heavily require a large amount of labeled data to learn an appropriate distance strategy, which limits the generalization and flexibility of the model. This study proposes a knowledge-enhanced hierarchical heterogeneous graph model, which adopts a global multi-view graph node encoding to acquire comprehensive personality features and their inherent associations, where three types of knowledge including part-of-speech (POS) tag, entity, and Linguistic Inquiry and Word Count (LIWC) are introduced. Then, a hierarchical heterogeneous graph with a “post-word-diverse knowledge” structure is constructed for each post to obtain enhanced representation. Finally, a relation guided representation optimization that considers intra-user relationships and inter-label relationships is further developed to learn more discriminative semantic representation. Experimental results on three widely used datasets demonstrate that the model outperforms state-of-the-art methods when training with only 100 samples (approximately 1% of the total data set).

Kuiyuan Zhang, Zhongyun Hua, Rushi Lan, Yifang Guo, Yushu Zhang, Guoai Xu

As deep learning techniques advance rapidly, deepfake speech synthesized through text-to-speech or voice conversion networks is becoming increasingly realistic, posing significant challenges for detection and raising potential threats to social security. This growing realism has prompted extensive research in speech deepfake detection. However, current detection methods primarily focus on extracting features from either the raw waveform or the spectrogram, often overlooking the valuable correspondences between these two modalities that could enhance the detection of previously unseen types of deepfakes. In this work, we propose a multi-view collaborative learning network for speech deepfake detection, which jointly learns robust speech representations from both raw waveforms and spectrograms. Specifically, we first design a Dual-Branch Contrastive Learning (DBCL) framework for learning different view features. DBCL consists of two branches that learn representations from the raw waveform or the spectrogram and utilizes contrastive learning to enhance inter- and inner-view correlations. Additionally, we introduce a Waveform-Spectrogram Fusion Module (WSFM) to exchange multi-view information for collaborative learning. In the feature learning process, WSFM converts features between views and merges them adaptively using waveform-spectrogram cross-attention. The final detection is conducted based on the concatenation of the waveform and spectrogram features. We conduct extensive experiments on four benchmark deepfake speech detection datasets, and the experimental results demonstrate that our method can achieve better detection performance than current state-of-the-art detection methods.