arXiv subjects

Pengyuan Zhang

Publications and source records attributed to Pengyuan Zhang.

At least 19 recordsLinked to original sources

arXiv2026

Beyond Binary Success: A Diagnostic Meta-Evaluation Framework for Fine-Grained Manipulation

Fine-grained manipulation marks a regime where global scene context no longer suffices, and success hinges on the tight coupling of local attribute grounding, high-fidelity spatial perception, and constraint-respecting motor execution. However, current embodied AI benchmarks collapse these capacities into binary success rates, systematically inflating reported capabilities by up to 70% and masking the architectural bottlenecks that impede real-world deployment. We introduce MetaFine, a diagnostic meta-evaluation framework that disentangles manipulation competency along three axes: understanding, perception, and controlled behavior. Built on a compositional task graph, MetaFine absorbs heterogeneous external benchmarks and reconstructs them into diagnostic scenarios of varying complexity under a unified protocol. Evaluating state-of-the-art vision-language-action (VLA) models through this lens exposes severe dimension-specific failures invisible to conventional metrics. Through targeted causal intervention, we identify the visual encoder's ability to preserve local spatial structure as a key bottleneck for fine-grained precision: improving it directly unlocks previously inaccessible manipulation capabilities without modifying downstream policies. MetaFine further supports hybrid real-sim validation, using limited paired real-world rollouts to calibrate scalable simulation-based estimates for more stable physical benchmarking. By shifting evaluation from ranking to diagnosis, MetaFine turns benchmarking into an actionable compass for repairing the layered capacities underlying genuine physical dexterity. The MetaFine framework, benchmarks, and supporting resources will be publicly released at our project page: https://metafine.github.io/.

cs.RO↗

arXiv2025

Emilia: A Large-Scale, Extensive, Multilingual, and Diverse Dataset for Speech Generation

Recent advancements in speech generation have been driven by large-scale training datasets. However, current models struggle to capture the spontaneity and variability inherent in real-world human speech, as they are primarily trained on audio-book datasets limited to formal, read-aloud speaking styles. To address this limitation, we introduce Emilia-Pipe, an open-source preprocessing pipeline designed to extract high-quality training data from valuable yet under-explored in-the-wild sources that capture spontaneous human speech in real-world contexts. Using Emilia-Pipe, we construct Emilia, which comprises over 101k hours of speech across six languages: English, Chinese, German, French, Japanese, and Korean. Furthermore, we expand Emilia to Emilia-Large, a dataset exceeding 216k hours, making it one of the largest open-source speech generation resources available. Extensive experiments show that Emilia-trained models produce markedly more spontaneous, human-like speech than those trained on traditional audio-book datasets, while matching their intelligibility. These models better capture diverse speaker timbres and the full spectrum of real-world conversational styles. Our work also highlights the importance of scaling dataset size for advancing speech generation performance and validates the effectiveness of Emilia for both multilingual and crosslingual speech generation tasks.

cs.SD↗

arXiv2025

SF-Speech: Straightened Flow for Zero-Shot Voice Clone

Recently, neural ordinary differential equations (ODE) models trained with flow matching have achieved impressive performance on the zero-shot voice clone task. Nevertheless, postulating standard Gaussian noise as the initial distribution of ODE gives rise to numerous intersections within the fitted targets of flow matching, which presents challenges to model training and enhances the curvature of the learned generated trajectories. These curved trajectories restrict the capacity of ODE models for generating desirable samples with a few steps. This paper proposes SF-Speech, a novel voice clone model based on ODE and in-context learning. Unlike the previous works, SF-Speech adopts a lightweight multi-stage module to generate a more deterministic initial distribution for ODE. Without introducing any additional loss function, we effectively straighten the curved reverse trajectories of the ODE model by jointly training it with the proposed module. Experiment results on datasets of various scales show that SF-Speech outperforms the state-of-the-art zero-shot TTS methods and requires only a quarter of the solver steps, resulting in a generation speed approximately 3.7 times that of Voicebox and E2 TTS. Audio samples are available at the demo page\footnote{[Online] Available: https://lixuyuan102.github.io/Demo/}.

cs.SD↗

arXiv2025

Controlling your Attributes in Voice

Attribute control in generative tasks aims to modify personal attributes, such as age and gender while preserving the identity information in the source sample. Although significant progress has been made in controlling facial attributes in image generation, similar approaches for speech generation remain largely unexplored. This letter proposes a novel method for controlling speaker attributes in speech without parallel data. Our approach consists of two main components: a GAN-based speaker representation variational autoencoder that extracts speaker identity and attributes from speaker vector, and a two-stage voice conversion model that captures the natural expression of speaker attributes in speech. Experimental results show that our proposed method not only achieves attribute control at the speaker representation level but also enables manipulation of the speaker age and gender at the speech level while preserving speech quality and speaker identity.

cs.SD↗

arXiv2025

SLIDE: Integrating Speech Language Model with LLM for Spontaneous Spoken Dialogue Generation

Recently, ``textless" speech language models (SLMs) based on speech units have made huge progress in generating naturalistic speech, including non-verbal vocalizations. However, the generated speech samples often lack semantic coherence. In this paper, we propose SLM and LLM Integration for spontaneous spoken Dialogue gEneration (SLIDE). Specifically, we first utilize an LLM to generate the textual content of spoken dialogue. Next, we convert the textual dialogues into phoneme sequences and use a two-tower transformer-based duration predictor to predict the duration of each phoneme. Finally, an SLM conditioned on the spoken phoneme sequences is used to vocalize the textual dialogue. Experimental results on the Fisher dataset demonstrate that our system can generate naturalistic spoken dialogue while maintaining high semantic coherence.

eess.AS↗

arXiv2024

Transliterated Zero-Shot Domain Adaptation for Automatic Speech Recognition

The performance of automatic speech recognition models often degenerates on domains not covered by the training data. Domain adaptation can address this issue, assuming the availability of the target domain data in the target language. However, such assumption does not stand in many real-world applications. To make domain adaptation more applicable, we address the problem of zero-shot domain adaptation (ZSDA), where target domain data is unavailable in the target language. Instead, we transfer the target domain knowledge from another source language where the target domain data is more accessible. To do that, we first perform cross-lingual pre-training (XLPT) to share domain knowledge across languages, then use target language fine-tuning to build the final model. One challenge in this practice is that the pre-trained knowledge can be forgotten during fine-tuning, resulting in sub-optimal adaptation performance. To address this issue, we propose transliterated ZSDA to achieve consistent pre-training and fine-tuning labels, leading to maximum preservation of the pre-trained knowledge. Experimental results show that transliterated ZSDA relatively decreases the word error rate by 9.2% compared with a wav2vec 2.0 baseline. Moreover, transliterated ZSDA consistently outperforms self-supervised ZSDA and performs on par with supervised ZSDA, proving the superiority of transliteration-based pre-training labels.

eess.AS↗

arXiv2024

The Database and Benchmark for the Source Speaker Tracing Challenge 2024

Voice conversion (VC) systems can transform audio to mimic another speaker's voice, thereby attacking speaker verification (SV) systems. However, ongoing studies on source speaker verification (SSV) are hindered by limited data availability and methodological constraints. This paper presents the Source Speaker Tracking Challenge (SSTC) on STL 2024, which aims to fill the gap in the database and benchmark for the SSV task. In this study, we generate a large-scale converted speech database with 16 common VC methods and train a batch of baseline systems based on the MFA-Conformer architecture. In addition, we introduced a related task called conversion method recognition, with the aim of assisting the SSV task. We expect SSTC to be a platform for advancing the development of the SSV task and provide further insights into the performance and limitations of current SV systems against VC attacks. Further details about SSTC can be found in https://sstc-challenge.github.io/.

eess.AS↗

arXiv2024

Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation

Recent advancements in speech generation models have been significantly driven by the use of large-scale training data. However, producing highly spontaneous, human-like speech remains a challenge due to the scarcity of large, diverse, and spontaneous speech datasets. In response, we introduce Emilia, the first large-scale, multilingual, and diverse speech generation dataset. Emilia starts with over 101k hours of speech across six languages, covering a wide range of speaking styles to enable more natural and spontaneous speech generation. To facilitate the scale-up of Emilia, we also present Emilia-Pipe, the first open-source preprocessing pipeline designed to efficiently transform raw, in-the-wild speech data into high-quality training data with speech annotations. Experimental results demonstrate the effectiveness of both Emilia and Emilia-Pipe. Demos are available at: https://emilia-dataset.github.io/Emilia-Demo-Page/.

eess.AS↗

arXiv2024

TRNet: Two-level Refinement Network leveraging Speech Enhancement for Noise Robust Speech Emotion Recognition

One persistent challenge in Speech Emotion Recognition (SER) is the ubiquitous environmental noise, which frequently results in deteriorating SER performance in practice. In this paper, we introduce a Two-level Refinement Network, dubbed TRNet, to address this challenge. Specifically, a pre-trained speech enhancement module is employed for front-end noise reduction and noise level estimation. Later, we utilize clean speech spectrograms and their corresponding deep representations as reference signals to refine the spectrogram distortion and representation shift of enhanced speech during model training. Experimental results validate that the proposed TRNet substantially promotes the robustness of the proposed system in both matched and unmatched noisy environments, without compromising its performance in noise-free environments.

cs.SD↗

arXiv2024

Expressive paragraph text-to-speech synthesis with multi-step variational autoencoder

Neural networks have been able to generate high-quality single-sentence speech. However, it remains a challenge concerning audio-book speech synthesis due to the intra-paragraph correlation of semantic and acoustic features as well as variable styles. In this paper, we propose a highly expressive paragraph speech synthesis system with a multi-step variational autoencoder, called EP-MSTTS. EP-MSTTS is the first VITS-based paragraph speech synthesis model and models the variable style of paragraph speech at five levels: frame, phoneme, word, sentence, and paragraph. We also propose a series of improvements to enhance the performance of this hierarchical model. In addition, we directly train EP-MSTTS on speech sliced by paragraph rather than sentence. Experiment results on the single-speaker French audiobook corpus released at Blizzard Challenge 2023 show EP-MSTTS obtains better performance than baseline models.

cs.SD↗

arXiv2024

Improving Short Utterance Anti-Spoofing with AASIST2

The wav2vec 2.0 and integrated spectro-temporal graph attention network (AASIST) based countermeasure achieves great performance in speech anti-spoofing. However, current spoof speech detection systems have fixed training and evaluation durations, while the performance degrades significantly during short utterance evaluation. To solve this problem, AASIST can be improved to AASIST2 by modifying the residual blocks to Res2Net blocks. The modified Res2Net blocks can extract multi-scale features and improve the detection performance for speech of different durations, thus improving the short utterance evaluation performance. On the other hand, adaptive large margin fine-tuning (ALMFT) has achieved performance improvement in short utterance speaker verification. Therefore, we apply Dynamic Chunk Size (DCS) and ALMFT training strategies in speech anti-spoofing to further improve the performance of short utterance evaluation. Experiments demonstrate that the proposed AASIST2 improves the performance of short utterance evaluation while maintaining the performance of regular evaluation on different datasets.

eess.AS↗

arXiv2023

Modality-Collaborative Transformer with Hybrid Feature Reconstruction for Robust Emotion Recognition

As a vital aspect of affective computing, Multimodal Emotion Recognition has been an active research area in the multimedia community. Despite recent progress, this field still confronts two major challenges in real-world applications: 1) improving the efficiency of constructing joint representations from unaligned multimodal features, and 2) relieving the performance decline caused by random modality feature missing. In this paper, we propose a unified framework, Modality-Collaborative Transformer with Hybrid Feature Reconstruction (MCT-HFR), to address these issues. The crucial component of MCT is a novel attention-based encoder which concurrently extracts and dynamically balances the intra- and inter-modality relations for all associated modalities. With additional modality-wise parameter sharing, a more compact representation can be encoded with less time and space complexity. To improve the robustness of MCT, we further introduce HFR which consists of two modules: Local Feature Imagination (LFI) and Global Feature Alignment (GFA). During model training, LFI leverages complete features as supervisory signals to recover local missing features, while GFA is designed to reduce the global semantic gap between pairwise complete and incomplete representations. Experimental evaluations on two popular benchmark datasets demonstrate that our proposed method consistently outperforms advanced baselines in both complete and incomplete data scenarios.

cs.CV↗

arXiv2023

DSNet: Disentangled Siamese Network with Neutral Calibration for Speech Emotion Recognition

One persistent challenge in deep learning based speech emotion recognition (SER) is the unconscious encoding of emotion-irrelevant factors (e.g., speaker or phonetic variability), which limits the generalization of SER in practical use. In this paper, we propose DSNet, a Disentangled Siamese Network with neutral calibration, to meet the demand for a more robust and explainable SER model. Specifically, we introduce an orthogonal feature disentanglement module to explicitly project the high-level representation into two distinct subspaces. Later, we propose a novel neutral calibration mechanism to encourage one subspace to capture sufficient emotion-irrelevant information. In this way, the other one can better isolate and emphasize the emotion-relevant information within speech signals. Experimental results on two popular benchmark datasets demonstrate the superiority of DSNet over various state-of-the-art methods for speaker-independent SER.

cs.SD↗

arXiv2023

Enhancing Spoofing Speech Detection Using Rhythm Information

Current spoofing speech detection systems need more convincing evidence. In this paper, the flaws of rhythm information inherent in the TTS-generated speech are analyzed to increase the reliability of detection systems. TTS models take text as input and utilize acoustic models to predict rhythm information, which introduces artifacts in the rhythm information. By filtering out vocal tract response, the remaining glottal flow with rhythm information retains detection ability for TTS-generated speech. Based on these analyses, a rhythm perturbation module is proposed to enhance the copy-synthesis data augmentation method. Fake utterances generated by the proposed method force the detecting model to pay attention to the artifacts in rhythm information and effectively improve the ability to detect TTS-generated speech of the anti-spoofing countermeasures.

eess.AS↗

arXiv2023

Synthetic Speech Detection Based on Temporal Consistency and Distribution of Speaker Features

Current synthetic speech detection (SSD) methods perform well on certain datasets but still face issues of robustness and interpretability. A possible reason is that these methods do not analyze the deficiencies of synthetic speech. In this paper, the flaws of the speaker features inherent in the text-to-speech (TTS) process are analyzed. Differences in the temporal consistency of intra-utterance speaker features arise due to the lack of fine-grained control over speaker features in TTS. Since the speaker representations in TTS are based on speaker embeddings extracted by encoders, the distribution of inter-utterance speaker features differs between synthetic and bonafide speech. Based on these analyzes, an SSD method based on temporal consistency and distribution of speaker features is proposed. On one hand, modeling the temporal consistency of intra-utterance speaker features can aid speech anti-spoofing. On the other hand, distribution differences in inter-utterance speaker features can be utilized for SSD. The proposed method offers low computational complexity and performs well in both cross-dataset and silence trimming scenarios.

eess.AS↗

arXiv2023

The Impact of Silence on Speech Anti-Spoofing

The current speech anti-spoofing countermeasures (CMs) show excellent performance on specific datasets. However, removing the silence of test speech through Voice Activity Detection (VAD) can severely degrade performance. In this paper, the impact of silence on speech anti-spoofing is analyzed. First, the reasons for the impact are explored, including the proportion of silence duration and the content of silence. The proportion of silence duration in spoof speech generated by text-to-speech (TTS) algorithms is lower than that in bonafide speech. And the content of silence generated by different waveform generators varies compared to bonafide speech. Then the impact of silence on model prediction is explored. Even after retraining, the spoof speech generated by neural network based end-to-end TTS algorithms suffers a significant rise in error rates when the silence is removed. To demonstrate the reasons for the impact of silence on CMs, the attention distribution of a CM is visualized through class activation mapping (CAM). Furthermore, the implementation and analysis of the experiments masking silence or non-silence demonstrates the significance of the proportion of silence duration for detecting TTS and the importance of silence content for detecting voice conversion (VC). Based on the experimental results, improving the robustness of CMs against unknown spoofing attacks by masking silence is also proposed. Finally, the attacks on anti-spoofing CMs through concatenating silence, and the mitigation of VAD and silence attack through low-pass filtering are introduced.

eess.AS↗

arXiv2023

One-Class Knowledge Distillation for Spoofing Speech Detection

The detection of spoofing speech generated by unseen algorithms remains an unresolved challenge. One reason for the lack of generalization ability is traditional detecting systems follow the binary classification paradigm, which inherently assumes the possession of prior knowledge of spoofing speech. One-class methods attempt to learn the distribution of bonafide speech and are inherently suited to the task where spoofing speech exhibits significant differences. However, training a one-class system using only bonafide speech is challenging. In this paper, we introduce a teacher-student framework to provide guidance for the training of a one-class model. The proposed one-class knowledge distillation method outperforms other state-of-the-art methods on the ASVspoof 21DF dataset and InTheWild dataset, which demonstrates its superior generalization ability.

eess.AS↗

arXiv2023

Alternative Pseudo-Labeling for Semi-Supervised Automatic Speech Recognition

When labeled data is insufficient, semi-supervised learning with the pseudo-labeling technique can significantly improve the performance of automatic speech recognition. However, pseudo-labels are often noisy, containing numerous incorrect tokens. Taking noisy labels as ground-truth in the loss function results in suboptimal performance. Previous works attempted to mitigate this issue by either filtering out the nosiest pseudo-labels or improving the overall quality of pseudo-labels. While these methods are effective to some extent, it is unrealistic to entirely eliminate incorrect tokens in pseudo-labels. In this work, we propose a novel framework named alternative pseudo-labeling to tackle the issue of noisy pseudo-labels from the perspective of the training objective. The framework comprises several components. Firstly, a generalized CTC loss function is introduced to handle noisy pseudo-labels by accepting alternative tokens in the positions of incorrect tokens. Applying this loss function in pseudo-labeling requires detecting incorrect tokens in the predicted pseudo-labels. In this work, we adopt a confidence-based error detection method that identifies the incorrect tokens by comparing their confidence scores with a given threshold, thus necessitating the confidence score to be discriminative. Hence, the second proposed technique is the contrastive CTC loss function that widens the confidence gap between the correctly and incorrectly predicted tokens, thereby improving the error detection ability. Additionally, obtaining satisfactory performance with confidence-based error detection typically requires extensive threshold tuning. Instead, we propose an automatic thresholding method that uses labeled data as a proxy for determining the threshold, thus saving the pain of manual tuning.

eess.AS↗