arXiv subjects

Zhepei Wang

Publications and source records attributed to Zhepei Wang.

At least 19 recordsLinked to original sources

arXiv2026

A Generative-First Neural Audio Autoencoder

Neural autoencoders underpin generative models. Practical, large-scale use of neural autoencoders for generative modeling necessitates fast encoding, low latent rates, and a single model across representations. Existing approaches are reconstruction-first: they incur high latent rates, slow encoding, and separate architectures for discrete vs. continuous latents and for different audio channel formats, hindering workflows from preprocessing to inference conditioning. We introduce a generative-first architecture for audio autoencoding that increases temporal downsampling from 2048x to 3360x and supports continuous and discrete representations and common audio channel formats in one model. By balancing compression, quality, and speed, it delivers 10x faster encoding, 1.6x lower rates, and eliminates channel-format-specific variants while maintaining competitive reconstruction quality. This enables applications previously constrained by processing costs: a 60-second mono signal compresses to 788 tokens, making generative modeling more tractable.

cs.SD↗

arXiv2026

TAC: Timestamped Audio Captioning

Large Audio Language Models struggle to disentangle overlapping events in complex acoustic scenes, yielding temporally inconsistent captions and frequent hallucinations. We introduce Timestamped Audio Captioner (TAC), a model that produces temporally grounded audio descriptions at varying degrees of detail and resolution. TAC is trained with a synthetic data pipeline that constructs challenging and dynamic mixtures from real-world audio sources, enabling robust learning under realistic polyphonic conditions. Across event detection and dense captioning, TAC outperforms all competing methods, with a low hallucination rate and accurate temporal grounding. We also introduce TAC-V, an audio-visual pipeline to generate semantically rich audio-visual descriptions. We then show that TAC and TAC-V serves as a "semantic bridge" for a text-only reasoner: a simple TAC$\rightarrow$LLM and TAC-V$\rightarrow$LLM cascade achieves state-of-the-art scores on benchmarks for both audio (MMAU-Pro, MMSU, MMAR) and audio-visual (DailyOmni, VideoHolmes) understanding and reasoning respectively.

cs.SD↗

arXiv2026

Rethinking Music Captioning with Music Metadata LLMs

Music captioning, or the task of generating a natural language description of music, is useful for both music understanding and controllable music generation. Training captioning models, however, typically requires high-quality music caption data which is scarce compared to metadata (e.g., genre, mood, etc.). As a result, it is common to use large language models (LLMs) to synthesize captions from metadata to generate training data for captioning models, though this process imposes a fixed stylization and entangles factual information with natural language style. As a more direct approach, we propose metadata-based captioning. We train a metadata prediction model to infer detailed music metadata from audio and then convert it into expressive captions via pre-trained LLMs at inference time. Compared to a strong end-to-end baseline trained on LLM-generated captions derived from metadata, our method: (1) achieves comparable performance in less training time over end-to-end captioners, (2) offers flexibility to easily change stylization post-training, enabling output captions to be tailored to specific stylistic and quality requirements, and (3) can be prompted with audio and partial metadata to enable powerful metadata imputation or in-filling--a common task for organizing music data.

cs.SD↗

arXiv2025

MultiPark: Multimodal Parking Transformer with Next-Segment Prediction

Parking accurately and safely in highly constrained spaces remains a critical challenge. Unlike structured driving environments, parking requires executing complex maneuvers such as frequent gear shifts and steering saturation. Recent attempts to employ imitation learning (IL) for parking have achieved promising results. However, existing works ignore the multimodal nature of parking behavior in lane-free open space, failing to derive multiple plausible solutions under the same situation. Notably, IL-based methods encompass inherent causal confusion, so enabling a neural network to generalize across diverse parking scenarios is particularly difficult. To address these challenges, we propose MultiPark, an autoregressive transformer for multimodal parking. To handle paths filled with abrupt turning points, we introduce a data-efficient next-segment prediction paradigm, enabling spatial generalization and temporal extrapolation. Furthermore, we design learnable parking queries factorized into gear, longitudinal, and lateral components, parallelly decoding diverse parking behaviors. To mitigate causal confusion in IL, our method employs target-centric pose and ego-centric collision as outcome-oriented loss across all modalities beyond pure imitation loss. Evaluations on real-world datasets demonstrate that MultiPark achieves state-of-the-art performance across various scenarios. We deploy MultiPark on a production vehicle, further confirming our approach's robustness in real-world parking environments.

cs.RO↗

arXiv2025

Fast Iterative Region Inflation for Computing Large 2-D/3-D Convex Regions of Obstacle-Free Space

Convex polytopes have compact representations and exhibit convexity, which makes them suitable for abstracting obstacle-free spaces from various environments. Existing generation methods struggle with balancing high-quality output and efficiency. Moreover, another crucial requirement for convex polytopes to accurately contain certain seed point sets, such as a robot or a front-end path, is proposed in various tasks, which we refer to as manageability. In this paper, we propose Fast Iterative Regional Inflation (FIRI) to generate high-quality convex polytope while ensuring efficiency and manageability simultaneously. FIRI consists of two iteratively executed submodules: Restrictive Inflation (RsI) and Maximum Volume Inscribed Ellipsoid (MVIE) computation. By explicitly incorporating constraints that include the seed point set, RsI guarantees manageability. Meanwhile, iterative MVIE optimization ensures high-quality result through monotonic volume bound improvement.In terms of efficiency, we design methods tailored to the low-dimensional and multi-constrained nature of both modules, resulting in orders of magnitude improvement compared to generic solvers. Notably, in 2-D MVIE, we present the first linear-complexity analytical algorithm for maximum area inscribed ellipse, further enhancing the performance in 2-D cases. Extensive benchmarks conducted against state-of-the-art methods validate the superior performance of FIRI in terms of quality, manageability, and efficiency. Furthermore, various real-world applications showcase the generality and practicality of FIRI.

cs.RO↗

arXiv2025

Primitive-Planner: An Ultra Lightweight Quadrotor Planner with Time-optimal Primitives

It is a significant requirement for a quadrotor trajectory planner to simultaneously guarantee trajectory quality and system lightweight. Many researchers focus on this problem, but there's still a gap between their performance and our common wish. In this paper, we propose an ultra lightweight quadrotor planner with time-optimal primitives. Firstly, a novel motion primitive library is proposed to generate time-optimal and dynamical feasible trajectories offline. Secondly, we propose a fast collision checking method with a deterministic time consumption, independent of the sampling resolution of the primitives. Finally, we select the minimum cost trajectory to execute among the safe primitives based on user-defined requirements. The propsed transformation relation between the local trajectories ensures the smoothness of the global trajectory. The planner reduces unnecessary online computing power consumption as much as possible, while ensuring a high-quality trajectory. Benchmark comparisons show that our method can generate the shortest flight time and distance of trajectory with the lowest computation overload. Challenging real-world experiments validate the robustness of our method.

cs.RO↗

arXiv2024

Audio Editing with Non-Rigid Text Prompts

In this paper, we explore audio-editing with non-rigid text edits. We show that the proposed editing pipeline is able to create audio edits that remain faithful to the input audio. We explore text prompts that perform addition, style transfer, and in-painting. We quantitatively and qualitatively show that the edits are able to obtain results which outperform Audio-LDM, a recently released text-prompted audio generation model. Qualitative inspection of the results points out that the edits given by our approach remain more faithful to the input audio in terms of keeping the original onsets and offsets of the audio events.

cs.SD↗

arXiv2024

On Class Separability Pitfalls In Audio-Text Contrastive Zero-Shot Learning

Recent advances in audio-text cross-modal contrastive learning have shown its potential towards zero-shot learning. One possibility for this is by projecting item embeddings from pre-trained backbone neural networks into a cross-modal space in which item similarity can be calculated in either domain. This process relies on a strong unimodal pre-training of the backbone networks, and on a data-intensive training task for the projectors. These two processes can be biased by unintentional data leakage, which can arise from using supervised learning in pre-training or from inadvertently training the cross-modal projection using labels from the zero-shot learning evaluation. In this study, we show that a significant part of the measured zero-shot learning accuracy is due to strengths inherited from the audio and text backbones, that is, they are not learned in the cross-modal domain and are not transferred from one modality to another.

cs.SD↗

arXiv2023

Unsupervised Improvement of Audio-Text Cross-Modal Representations

Recent advances in using language models to obtain cross-modal audio-text representations have overcome the limitations of conventional training approaches that use predefined labels. This has allowed the community to make progress in tasks like zero-shot classification, which would otherwise not be possible. However, learning such representations requires a large amount of human-annotated audio-text pairs. In this paper, we study unsupervised approaches to improve the learning framework of such representations with unpaired text and audio. We explore domain-unspecific and domain-specific curation methods to create audio-text pairs that we use to further improve the model. We also show that when domain-specific curation is used in conjunction with a soft-labeled contrastive loss, we are able to obtain significant improvement in terms of zero-shot classification performance on downstream sound event classification or acoustic scene classification tasks.

cs.SD↗

arXiv2023

A Framework for Unified Real-time Personalized and Non-Personalized Speech Enhancement

In this study, we present an approach to train a single speech enhancement network that can perform both personalized and non-personalized speech enhancement. This is achieved by incorporating a frame-wise conditioning input that specifies the type of enhancement output. To improve the quality of the enhanced output and mitigate oversuppression, we experiment with re-weighting frames by the presence or absence of speech activity and applying augmentations to speaker embeddings. By training under a multi-task learning setting, we empirically show that the proposed unified model obtains promising results on both personalized and non-personalized speech enhancement benchmarks and reaches similar performance to models that are trained specialized for either task. The strong performance of the proposed method demonstrates that the unified model is a more economical alternative compared to keeping separate task-specific models during inference.

eess.AS↗

arXiv2023

A Linear and Exact Algorithm for Whole-Body Collision Evaluation via Scale Optimization

Collision evaluation is of vital importance in various applications. However, existing methods are either cumbersome to calculate or have a gap with the actual value. In this paper, we propose a zero-gap whole-body collision evaluation which can be formulated as a low dimensional linear program. This evaluation can be solved analytically in O(m) computational time, where m is the total number of the linear inequalities in this linear program. Moreover, the proposed method is efficient in obtaining its gradient, making it easy to apply to optimization-based applications.

cs.RO↗

arXiv2022

Learning Representations for New Sound Classes With Continual Self-Supervised Learning

In this paper, we work on a sound recognition system that continually incorporates new sound classes. Our main goal is to develop a framework where the model can be updated without relying on labeled data. For this purpose, we propose adopting representation learning, where an encoder is trained using unlabeled data. This learning framework enables the study and implementation of a practically relevant use case where only a small amount of the labels is available in a continual learning context. We also make the empirical observation that a similarity-based representation learning method within this framework is robust to forgetting even if no explicit mechanism against forgetting is employed. We show that this approach obtains similar performance compared to several distillation-based continual learning methods when employed on self-supervised representation learning methods.

eess.AS↗

arXiv2022

Bubble Planner: Planning High-speed Smooth Quadrotor Trajectories using Receding Corridors

Quadrotors are agile platforms. With human experts, they can perform extremely high-speed flights in cluttered environments. However, fully autonomous flight at high speed remains a significant challenge. In this work, we propose a motion planning algorithm based on the corridor-constrained minimum control effort trajectory optimization (MINCO) framework. Specifically, we use a series of overlapping spheres to represent the free space of the environment and propose two novel designs that enable the algorithm to plan high-speed quadrotor trajectories in real-time. One is a sampling-based corridor generation method that generates spheres with large overlapped areas (hence overall corridor size) between two neighboring spheres. The second is a Receding Horizon Corridors (RHC) strategy, where part of the previously generated corridor is reused in each replan. Together, these two designs enlarge the corridor spaces in accordance with the quadrotor's current state and hence allow the quadrotor to maneuver at high speeds. We benchmark our algorithm against other state-of-the-art planning methods to show its superiority in simulation. Comprehensive ablation studies are also conducted to show the necessity of the two designs. The proposed method is finally evaluated on an autonomous LiDAR-navigated quadrotor UAV in woods environments, achieving flight speeds over 13.7 m/s without any prior map of the environment or external localization facility.

cs.RO↗

arXiv2022

Semi-supervised Time Domain Target Speaker Extraction with Attention

In this work, we propose Exformer, a time-domain architecture for target speaker extraction. It consists of a pre-trained speaker embedder network and a separator network based on transformer encoder blocks. We study multiple methods to combine speaker information with the input mixture, and the resulting Exformer architecture obtains superior extraction performance compared to prior time-domain networks. Furthermore, we investigate a two-stage procedure to train the model using mixtures without reference signals upon a pre-trained supervised model. Experimental results show that the proposed semi-supervised learning procedure improves the performance of the supervised baselines.

eess.AS↗

arXiv2022

Generating Large-Scale Trajectories Efficiently using Double Descriptions of Polynomials

For quadrotor trajectory planning, describing a polynomial trajectory through coefficients and end-derivatives both enjoy their own convenience in energy minimization. We name them double descriptions of polynomial trajectories. The transformation between them, causing most of the inefficiency and instability, is formally analyzed in this paper. Leveraging its analytic structure, we design a linear-complexity scheme for both jerk/snap minimization and parameter gradient evaluation, which possesses efficiency, stability, flexibility, and scalability. With the help of our scheme, generating an energy optimal (minimum snap) trajectory only costs 1 $μs$ per piece at the scale up to 1,000,000 pieces. Moreover, generating large-scale energy-time optimal trajectories is also accelerated by an order of magnitude against conventional methods.

cs.RO↗

arXiv2022

Geometrically Constrained Trajectory Optimization for Multicopters

We present an optimization-based framework for multicopter trajectory planning subject to geometrical configuration constraints and user-defined dynamic constraints. The basis of the framework is a novel trajectory representation built upon our novel optimality conditions for unconstrained control effort minimization. We design linear-complexity operations on this representation to conduct spatial-temporal deformation under various planning requirements. Smooth maps are utilized to exactly eliminate geometrical constraints in a lightweight fashion. A variety of state-input constraints are supported by the decoupling of dense constraint evaluation from sparse parameterization, and backward differentiation of flatness map. As a result, this framework transforms a generally constrained multicopter planning problem into an unconstrained optimization that can be solved reliably and efficiently. Our framework bridges the gaps among solution quality, planning efficiency, and constraint fidelity for a multicopter with limited resources and maneuvering capability. Its generality and robustness are both demonstrated by applications to different flight tasks. Extensive simulations and benchmarks are also conducted to show its capability of generating high-quality solutions while retaining the computation speed against other specialized methods by orders of magnitude. The source code of our framework is available at: https://github.com/ZJU-FAST-Lab/GCOPTER

cs.RO↗

arXiv2022

Star-Convex Constrained Optimization for Visibility Planning with Application to Aerial Inspection

The visible capability is critical in many robot applications, such as inspection and surveillance, etc. Without the assurance of the visibility to targets, some tasks end up not being complete or even failing. In this paper, we propose a visibility guaranteed planner by star-convex constrained optimization. The visible space is modeled as star convex polytope (SCP) by nature and is generated by finding the visible points directly on point cloud. By exploiting the properties of the SCP, the visibility constraint is formulated for trajectory optimization. The trajectory is confined in the safe and visible flight corridor which consists of convex polytopes and SCPs. We further make a relaxation to the visibility constraints and transform the constrained trajectory optimization problem into an unconstrained one that can be reliably and efficiently solved. To validate the capability of the proposed planner, we present the practical application in site inspection. The experimental results show that the method is efficient, scalable, and visibility guaranteed, presenting the prospect of application to various other applications in the future.

cs.RO↗

arXiv2022

Improved singing voice separation with chromagram-based pitch-aware remixing

Singing voice separation aims to separate music into vocals and accompaniment components. One of the major constraints for the task is the limited amount of training data with separated vocals. Data augmentation techniques such as random source mixing have been shown to make better use of existing data and mildly improve model performance. We propose a novel data augmentation technique, chromagram-based pitch-aware remixing, where music segments with high pitch alignment are mixed. By performing controlled experiments in both supervised and semi-supervised settings, we demonstrate that training models with pitch-aware remixing significantly improves the test signal-to-distortion ratio (SDR)

eess.AS↗