Can we determine someone's geographic location purely from the sounds they hear? Are acoustic signals enough to localize within a country, state, or even city? We tackle the challenge of global-scale audio geolocation, formalize the problem, and conduct an in-depth analysis with wildlife audio from the iNatSounds dataset. Adopting a vision-inspired approach, we convert audio recordings to spectrograms and benchmark existing image geolocation techniques. We hypothesize that species vocalizations offer strong geolocation cues due to their defined geographic ranges and propose an approach that integrates species range prediction with retrieval-based geolocation. We further evaluate whether geolocation improves when analyzing species-rich recordings or when aggregating across spatiotemporal neighborhoods. Finally, we introduce case studies from movies to explore multimodal geolocation using both audio and visual content. Our work highlights the advantages of integrating audio and visual cues, and sets the stage for future research in audio geolocation.
Speech-based Parkinson's disease (PD) detection has gained attention for its automated, cost-effective, and non-intrusive nature. As research studies usually rely on data from diagnostic-oriented speech tasks, this work explores the feasibility of diagnosing PD on the basis of speech data not originally intended for diagnostic purposes, using the Turn-Taking (TT) dataset. Our findings indicate that TT can be as useful as diagnostic-oriented PD datasets like PC-GITA. We also investigate which specific dataset characteristics impact PD classification performance. The results show that concatenating audio recordings and balancing participants' gender and status distributions can be beneficial. Cross-dataset evaluation reveals that models trained on PC-GITA generalize poorly to TT, whereas models trained on TT perform better on PC-GITA. Furthermore, we provide insights into the high variability across folds, which is mainly due to large differences in individual speaker performance.
Large Language Models (LLMs) have demonstrated effectiveness in code generation tasks. To enable LLMs to address more complex coding challenges, existing research has focused on crafting multi-agent systems with agentic workflows, where complex coding tasks are decomposed into sub-tasks, assigned to specialized agents. Despite their effectiveness, current approaches heavily rely on hand-crafted agentic workflows, with both agent topologies and prompts manually designed, which limits their ability to automatically adapt to different types of coding problems. To address these limitations and enable automated workflow design, we propose \textbf{S}elf-\textbf{E}volving \textbf{W}orkflow (\textbf{SEW}), a novel self-evolving framework that automatically generates and optimises multi-agent workflows. Extensive experiments on three coding benchmark datasets, including the challenging LiveCodeBench, demonstrate that our SEW can automatically design agentic workflows and optimise them through self-evolution, bringing up to 33\% improvement on LiveCodeBench compared to using the backbone LLM only. Furthermore, by investigating different representation schemes of workflow, we provide insights into the optimal way to encode workflow information with text.
Speech emotion recognition (SER) in naturalistic settings remains a challenge due to the intrinsic variability, diverse recording conditions, and class imbalance. As participants in the Interspeech Naturalistic SER Challenge which focused on these complexities, we present Abhinaya, a system integrating speech-based, text-based, and speech-text models. Our approach fine-tunes self-supervised and speech large language models (SLLM) for speech representations, leverages large language models (LLM) for textual context, and employs speech-text modeling with an SLLM to capture nuanced emotional cues. To combat class imbalance, we apply tailored loss functions and generate categorical decisions through majority voting. Despite one model not being fully trained, the Abhinaya system ranked 4th among 166 submissions. Upon completion of training, it achieved state-of-the-art performance among published results, demonstrating the effectiveness of our approach for SER in real-world conditions.
Spatial hearing, the brain's ability to use auditory cues to identify the origin of sounds, is crucial for everyday listening. While simplified paradigms have advanced the understanding of spatial hearing, their lack of ecological validity limits their applicability to real-life conditions. This study aims to address this gap by investigating the effects of listener movement, reverberation, and distance on localisation accuracy in a more ecologically valid context. Participants performed active localisation tasks with no specific instructions on listening strategy, in either anechoic or reverberant conditions. The results indicate that the head movements were more frequent in reverberant environments, suggesting an adaptive strategy to mitigate uncertainty in binaural cues due to reverberation. While distance did not affect the listening strategy, it influenced the localisation performance. Our outcomes suggest that listening behaviour is adapted depending on the current acoustic conditions to support an effective perception of the space.
In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.
Speaker anonymization seeks to conceal a speaker's identity while preserving the utility of their speech. The achieved privacy is commonly evaluated with a speaker recognition model trained on anonymized speech. Although this represents a strong attack, it is unclear which aspects of speech are exploited to identify the speakers. Our research sets out to unveil these aspects. It starts with kNN-VC, a powerful voice conversion model that performs poorly as an anonymization system, presumably because of prosody leakage. To test this hypothesis, we extend kNN-VC with two interpretable components that anonymize the duration and variation of phones. These components increase privacy significantly, proving that the studied prosodic factors encode speaker identity and are exploited by the privacy attack. Additionally, we show that changes in the target selection algorithm considerably influence the outcome of the privacy attack.
The rapid growth of voice assistants powered by large language models (LLM) has highlighted a need for speech instruction data to train these systems. Despite the abundance of speech recognition data, there is a notable scarcity of speech instruction data, which is essential for fine-tuning models to understand and execute spoken commands. Generating high-quality synthetic speech requires a good text-to-speech (TTS) model, which may not be available to low resource languages. Our novel approach addresses this challenge by halting synthesis at the semantic representation level, bypassing the need for TTS. We achieve this by aligning synthetic semantic representations with the pre-trained Whisper encoder, enabling an LLM to be fine-tuned on text instructions while maintaining the ability to understand spoken instructions during inference. This simplified training process is a promising approach to building voice assistant for low-resource languages.
Generative error correction (GER) with large language models (LLMs) has emerged as an effective post-processing approach to improve automatic speech recognition (ASR) performance. However, it often struggles with rare or domain-specific words due to limited training data. Furthermore, existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction. To address these issues, we propose a novel LLM-based GER approach that targets rare words and incorporates phonetic information. First, we generate synthetic data to contain rare words for fine-tuning the GER model. Second, we integrate ASR's N-best hypotheses along with phonetic context to mitigate over-correction. Experimental results show that our method not only improves the correction of rare words but also reduces the WER and CER across both English and Japanese datasets.
Music auto-tagging is essential for organizing and discovering music in extensive digital libraries. While foundation models achieve exceptional performance in this domain, their outputs often lack interpretability, limiting trust and usability for researchers and end-users alike. In this work, we present an interpretable framework for music auto-tagging that leverages groups of musically meaningful multimodal features, derived from signal processing, deep learning, ontology engineering, and natural language processing. To enhance interpretability, we cluster features semantically and employ an expectation maximization algorithm, assigning distinct weights to each group based on its contribution to the tagging process. Our method achieves competitive tagging performance while offering a deeper understanding of the decision-making process, paving the way for more transparent and user-centric music tagging systems.
Generating high-quality piano audio from video requires precise synchronization between visual cues and musical output, ensuring accurate semantic and temporal alignment.However, existing evaluation datasets do not fully capture the intricate synchronization required for piano music generation. A comprehensive benchmark is essential for two primary reasons: (1) existing metrics fail to reflect the complexity of video-to-piano music interactions, and (2) a dedicated benchmark dataset can provide valuable insights to accelerate progress in high-quality piano music generation. To address these challenges, we introduce the CoP Benchmark Dataset-a fully open-sourced, multimodal benchmark designed specifically for video-guided piano music generation. The proposed Chain-of-Perform (CoP) benchmark offers several compelling features: (1) detailed multimodal annotations, enabling precise semantic and temporal alignment between video content and piano audio via step-by-step Chain-of-Perform guidance; (2) a versatile evaluation framework for rigorous assessment of both general-purpose and specialized video-to-piano generation tasks; and (3) full open-sourcing of the dataset, annotations, and evaluation protocols. The dataset is publicly available at https://github.com/acappemin/Video-to-Audio-and-Piano, with a continuously updated leaderboard to promote ongoing research in this domain.
Emotion plays a fundamental role in human interaction, and therefore systems capable of identifying emotions in speech are crucial in the context of human-computer interaction. Speech emotion recognition (SER) is a challenging problem, particularly in natural speech and when the available data is imbalanced across emotions. This paper presents our proposed system in the context of the 2025 Speech Emotion Recognition in Naturalistic Conditions Challenge. Our proposed architecture leverages cross-modality, utilizing cross-modal attention to fuse representations from different modalities. To address class imbalance, we employed two training designs: (i) weighted crossentropy loss (WCE); and (ii) WCE with an additional neutralexpressive soft margin loss and balancing. We trained a total of 12 multimodal models, which were ensembled using a balanced stacking model. Our proposed system achieves a MacroF1 score of 0.4094 and an accuracy of 0.4128 on 8-class speech emotion recognition.
Due to the powerful vision-language reasoning and generalization abilities, multimodal large language models (MLLMs) have garnered significant attention in the field of end-to-end (E2E) autonomous driving. However, their application to closed-loop systems remains underexplored, and current MLLM-based methods have not shown clear superiority to mainstream E2E imitation learning approaches. In this work, we propose ReasonPlan, a novel MLLM fine-tuning framework designed for closed-loop driving through holistic reasoning with a self-supervised Next Scene Prediction task and supervised Decision Chain-of-Thought process. This dual mechanism encourages the model to align visual representations with actionable driving context, while promoting interpretable and causally grounded decision making. We curate a planning-oriented decision reasoning dataset, namely PDR, comprising 210k diverse and high-quality samples. Our method outperforms the mainstream E2E imitation learning method by a large margin of 19% L2 and 16.1 driving score on Bench2Drive benchmark. Furthermore, ReasonPlan demonstrates strong zero-shot generalization on unseen DOS benchmark, highlighting its adaptability in handling zero-shot corner cases. Code and dataset will be found in https://github.com/Liuxueyi/ReasonPlan.
Sparse-reward reinforcement learning (RL) can model a wide range of highly complex tasks. Solving sparse-reward tasks is RL's core premise - requiring efficient exploration coupled with long-horizon credit assignment - and overcoming these challenges is key for building self-improving agents with superhuman ability. We argue that solving complex and high-dimensional tasks requires solving simpler tasks that are relevant to the target task. In contrast, most prior work designs strategies for selecting exploratory tasks with the objective of solving any task, making exploration of challenging high-dimensional, long-horizon tasks intractable. We find that the sense of direction, necessary for effective exploration, can be extracted from existing RL algorithms, without needing any prior information. Based on this finding, we propose a method for directed sparse-reward goal-conditioned very long-horizon RL (DISCOVER), which selects exploratory goals in the direction of the target task. We connect DISCOVER to principled exploration in bandits, formally bounding the time until the target task becomes achievable in terms of the agent's initial distance to the target, but independent of the volume of the space of all tasks. Empirically, we perform a thorough evaluation in high-dimensional environments. We find that the directed goal selection of DISCOVER solves exploration problems that are beyond the reach of prior state-of-the-art exploration methods in RL.
Learning unknown dynamics under environmental (or external) constraints is fundamental to many fields (e.g., modern robotics), particularly challenging when constraint information is only locally available and uncertain. Existing approaches requiring global constraints or using probabilistic filtering fail to fully exploit the geometric structure inherent in local measurements (by using, e.g., sensors) and constraints. This paper presents a geometric framework unifying measurements, constraints, and dynamics learning through a fiber bundle structure over the state space. This naturally induced geometric structure enables measurement-aware Control Barrier Functions that adapt to local sensing (or measurement) conditions. By integrating Neural ODEs, our framework learns continuous-time dynamics while preserving geometric constraints, with theoretical guarantees of learning convergence and constraint satisfaction dependent on sensing quality. The geometric framework not only enables efficient dynamics learning but also suggests promising directions for integration with reinforcement learning approaches. Extensive simulations demonstrate significant improvements in both learning efficiency and constraint satisfaction over traditional methods, especially under limited and uncertain sensing conditions.
To address the gaps between the static pre-set "thinking-planning-action" of humanoid robots in unfamiliar scenarios and the highly programmed "call tool-return result" due to the lack of autonomous coding capabilities, this work designs a dynamic architecture connecting continuous thought machines (CTM) and model context protocol (MCP). It proposes a theoretical parallel solution through tick-slab and uses rank compression to achieve parameter suppression to provide a solution for achieving autonomous actions due to autonomous coding. The researcher used a simulation-based experiment using OpenAI's o4-mini-high as a tool to build the experimental environment, and introduced the extended SayCan dataset to conduct nine epochs of experiments. The experimental results show that the CTM-MCP architecture is feasible and effective through the data results of seven metrics: task success rate (TSR), execution success rate (ESR), average episode length (AEL), ROSCOE, REVEAL, proficiency self-assessment (PSA), task effectiveness (TE). In practice, it provides a reference experience for exploring the autonomous dynamic coding of humanoid robots based on continuous thinking to achieve human-like autonomous actions.
Modern policy gradient algorithms, such as TRPO and PPO, outperform vanilla policy gradient in many RL tasks. Questioning the common belief that enforcing approximate trust regions leads to steady policy improvement in practice, we show that the more critical factor is the enhanced value estimation accuracy from more value update steps in each iteration. To demonstrate, we show that by simply increasing the number of value update steps per iteration, vanilla policy gradient itself can achieve performance comparable to or better than PPO in all the standard continuous control benchmark environments. Importantly, this simple change to vanilla policy gradient is significantly more robust to hyperparameter choices, opening up the possibility that RL algorithms may still become more effective and easier to use.
Imitation from videos often fails when expert demonstrations and learner environments exhibit domain shifts, such as discrepancies in lighting, color, or texture. While visual randomization partially addresses this problem by augmenting training data, it remains computationally intensive and inherently reactive, struggling with unseen scenarios. We propose a different approach: instead of randomizing appearances, we eliminate their influence entirely by rethinking the sensory representation itself. Inspired by biological vision systems that prioritize temporal transients (e.g., retinal ganglion cells) and by recent sensor advancements, we introduce event-inspired perception for visually robust imitation. Our method converts standard RGB videos into a sparse, event-based representation that encodes temporal intensity gradients, discarding static appearance features. This biologically grounded approach disentangles motion dynamics from visual style, enabling robust visual imitation from observations even in the presence of visual mismatches between expert and agent environments. By training policies on event streams, we achieve invariance to appearance-based distractors without requiring computationally expensive and environment-specific data augmentation techniques. Experiments across the DeepMind Control Suite and the Adroit platform for dynamic dexterous manipulation show the efficacy of our method. Our code is publicly available at Eb-LAIfO.
Recent high-capacity vision-language-action (VLA) models have demonstrated impressive performance on a range of robotic manipulation tasks by imitating human demonstrations. However, exploiting offline data with limited visited states will cause execution failure in out-of-distribution scenarios. Intuitively, an exploration-based method that improves on online collected data at test time could address this limitation. We present VLA-RL, an algorithmic and systematic framework that leverages online reinforcement learning (RL) to improve pretrained auto-regressive VLAs in downstream tasks. Within a unified perspective, we first introduce a trajectory-level RL formulation for auto-regressive VLA training, which models general robotic manipulation trajectory as multi-modal multi-turn conversation. To address the challenge of sparse rewards, we fine-tune a pretrained vision-language model as a robotic process reward model, which is trained on pseudo reward labels annotated on automatically extracted task segments. To scale up, we identify several implementation findings that improve the stability and efficiency including curriculum selection strategy, GPU-balanced vectorized environments, batch decoding, and critic warmup. VLA-RL enables OpenVLA-7B to surpass the strongest finetuned baseline by 4.5% on 40 challenging robotic manipulation tasks in LIBERO, and even matches the performance of advanced commercial models such as $\pi_0$-FAST. Notably, we observe that VLA-RL benefits from increased test-time optimization, indicating an early spark of inference scaling laws in robotics.
Ballbot (i.e. Ball balancing robot) navigation usually relies on methods rooted in control theory (CT), and works that apply Reinforcement learning (RL) to the problem remain rare while generally being limited to specific subtasks (e.g. balance recovery). Unlike CT based methods, RL does not require (simplifying) assumptions about environment dynamics (e.g. the absence of slippage between the ball and the floor). In addition to this increased accuracy in modeling, RL agents can easily be conditioned on additional observations such as depth-maps without the need for explicit formulations from first principles, leading to increased adaptivity. Despite those advantages, there has been little to no investigation into the capabilities, data-efficiency and limitations of RL based methods for ballbot control and navigation. Furthermore, there is a notable absence of an open-source, RL-friendly simulator for this task. In this paper, we present an open-source ballbot simulation based on MuJoCo, and show that with appropriate conditioning on exteroceptive observations as well as reward shaping, policies learned by classical model-free RL methods are capable of effectively navigating through randomly generated uneven terrain, using a reasonable amount of data (four to five hours on a system operating at 500hz).