In this work, we propose Mutual Forcing, a framework for fast autoregressive audio-video generation with long-horizon audio-video synchronization. Our approach addresses two key challenges: joint audio-video modeling and fast autoregressive generation. To ease joint audio-video optimization, we adopt a two-stage training strategy: we first train uni-modal generators and then couple them into a unified audio-video model for joint training on paired data. For streaming generation, we ask whether a native fast causal audio-video model can be trained directly, instead of following existing streaming distillation pipelines that typically train a bidirectional model first and then convert it into a causal generator through multiple distillation stages. Our answer is Mutual Forcing, which builds directly on native autoregressive model and integrates few-step and multi-step generation within a single weight-shared model, enabling self-distillation and improved training-inference consistency. The multi-step mode improves the few-step mode via self-distillation, while the few-step mode generates historical context during training to improve training-inference consistency; because the two modes share parameters, these two effects reinforce each other within a single model. Compared with prior approaches such as Self-Forcing, Mutual Forcing removes the need for an additional bidirectional teacher model, supports more flexible training sequence lengths, reduces training overhead, and allows the model to improve directly from real paired data rather than a fixed teacher. Experiments show that Mutual Forcing matches or surpasses strong baselines that require around 50 sampling steps while using only 4 to 8 steps, demonstrating substantial advantages in both efficiency and quality. The project page is available at https://mutualforcing.github.io.

Real-time automatic speech recognition (ASR) systems face a fundamental trade-off between transcription accuracy and computational efficiency, particularly when deploying large-scale transformer models like Whisper. Existing streaming approaches either sacrifice accuracy through aggressive chunking or incur prohibitive memory costs through unbounded context accumulation. We present WhisperPipe, a novel streaming architecture that achieves bounded memory consumption while maintaining transcription quality through three key innovations a hybrid Voice Activity Detection (VAD) pipeline combining Silero VAD with energy-based filtering to reduce false activations by 34%, a dynamic buffering mechanism with overlapping context windows that prevents information loss at segment boundaries, and an adaptive processing strategy that balances latency and accuracy based on speech characteristics. Evaluated on 2.5 hours of diverse audio data, WhisperPipe demonstrates a median end-to-end latency of 89ms (90th percentile: 142ms) while consuming 48% less peak GPU memory and 80.9% lower average GPU utilization compared to baseline Whisper implementations. The system maintains stable memory usage over extended sessions, with zero growth rate across 150-minute continuous operation. Comparative analysis against related work shows that WhisperPipe achieves competitive accuracy (WER within 2% of offline Whisper) while operating at 3-5x lower latency than existing streaming solutions. The architecture's modular design enables deployment across resource-constrained environments, from edge devices to cloud infrastructure. Our results demonstrate that careful architectural design can reconcile the competing demands of real-time responsiveness and model sophistication in production ASR systems.

Recent audio-aware large language models (ALLMs) have demonstrated strong capabilities across diverse audio understanding and reasoning tasks, but they still frequently produce hallucinated or overly confident outputs. While uncertainty estimation has been extensively studied in text-only LLMs, it remains largely unexplored for ALLMs, where audio-conditioned generation introduces additional challenges such as perceptual ambiguity and cross-modal grounding. In this work, we present the first systematic empirical study of uncertainty estimation in ALLMs. We benchmark five representative methods, including predictive entropy, length-normalized entropy, semantic entropy, discrete semantic entropy, and P(True), across multiple models and diverse evaluation settings spanning general audio understanding, reasoning, hallucination detection, and unanswerable question answering. Our results reveal two key findings. First, semantic-level and verification-based methods consistently outperform token-level baselines on general audio reasoning benchmarks. Second, on trustworthiness-oriented benchmarks, the relative effectiveness of uncertainty methods becomes notably more model- and benchmark-dependent, indicating that conclusions drawn from general reasoning settings do not straightforwardly transfer to hallucination and unanswerable-question scenarios. We further explore uncertainty-based adaptive inference as a potential downstream application. We hope this study provides a foundation for future research on reliable, uncertainty-aware audio-language systems.

Voice assistants increasingly rely on Speech Language Models (SpeechLMs) to interpret spoken queries and execute complex tasks, yet existing benchmarks lack domain breadth, acoustic diversity, and compositional reasoning complexity to evaluate tool-calling performance. We introduce Audio2Tool, a large-scale dataset comprising approximately 30,000 queries designed to assess tool-calling capabilities of SpeechLMs across three primary domains: Smart Car, Smart Home, and Wearables. Our benchmark features a multi-tier complexity hierarchy, ranging from simple direct commands to complex multi-intent and needle-in-a-haystack extraction to isolate distinct failure modes. To ensure realism, we employ zero-shot voice cloning text-to-speech synthesis and diverse noise profiles to simulate in-the-wild conditions. Evaluations of state-of-the-art SpeechLMs and ASR-LLM pipelines show strong performance on simple commands but significant degradation under compositional and acoustic challenges. Code and dataset are publicly available on the project page: https://audio2tool.github.io/.

Recent advances in Speech Large Language Models (Speech-LLMs) have made significant progress, greatly enhancing multimodal interaction capabilities.However, their application in low-resource and dialect-diverse environments still faces challenges. The severe scarcity of Tibetan data, coupled with the phonetic differences among its major dialects (Ü-Tsang, Amdo, and Kham), is a prime example of this challenge. This paper proposes Ti-Audio, the first multi-dialectal end-to-end Speech-LLM for Tibetan. To efficiently align speech and text, we introduce a Dynamic Q-Former Adapter that extracts essential acoustic features from variable-length speech, ensuring stable cross-modal alignment even with limited data. At the data level, we leverage mutual assistance among related dialects to alleviate data scarcity and employ a temperature-based sampling strategy to maximize this synergy. Experimental results demonstrate that Ti-Audio achieves state-of-the-art performance on Tibetan benchmarks for automatic speech recognition and speech translation. Our work validates the effectiveness of cross-dialectal cooperation and provides a scalable paradigm for the development of Speech-LLM in low-resource scenarios.

Speaker embedding learning based on Euclidean space has achieved significant progress, but it is still insufficient in modeling hierarchical information within speaker features. Hyperbolic space, with its negative curvature geometric properties, can efficiently represent hierarchical information within a finite volume, making it more suitable for the feature distribution of speaker embeddings. In this paper, we propose Hyperbolic Softmax (H-Softmax) and Hyperbolic Additive Margin Softmax (HAM-Softmax) based on hyperbolic space. H-Softmax incorporates hierarchical information into speaker embeddings by projecting embeddings and speaker centers into hyperbolic space and computing hyperbolic distances. HAM-Softmax further enhances inter-class separability by introducing margin constraint on this basis. Experimental results show that H-Softmax and HAM-Softmax achieve average relative EER reductions of 27.84% and 14.23% compared with standard Softmax and AM-Softmax, respectively, demonstrating that the proposed methods effectively improve speaker verification performance and at the same time preserve the capability of hierarchical structure modeling. The code will be released at https://github.com/PunkMale/HAM-Softmax.

This paper introduces our solution, XM-ALIGN (Unified Cross-Modal Embedding Alignment Framework), proposed for the FAME challenge at ICASSP 2026. Our framework combines explicit and implicit alignment mechanisms, significantly improving cross-modal verification performance in both "heard" and "unheard" languages. By extracting feature embeddings from both face and voice encoders and jointly optimizing them using a shared classifier, we employ mean squared error (MSE) as the embedding alignment loss to ensure tight alignment between modalities. Additionally, data augmentation strategies are applied during model training to enhance generalization. Experimental results show that our approach demonstrates superior performance on the MAV-Celeb dataset. The code will be released at https://github.com/PunkMale/XM-ALIGN.

The deep learning models used for speaker verification rely heavily on large amounts of data and correct labeling. However, noisy (incorrect) labels often occur, which degrades the performance of the system. In this paper, we propose a novel two-stage learning method to filter out noisy labels from speaker datasets. Since a DNN will first fit data with clean labels, we first train the model with all data for several epochs. Then, based on this model, the model predictions are compared with the labels using our proposed the OR-Gate with top-k mechanism to select the data with clean labels and the selected data is used to train the model. This process is iterated until the training is completed. We have demonstrated the effectiveness of this method in filtering noisy labels through extensive experiments and have achieved excellent performance on the VoxCeleb (1 and 2) with different added noise rates.

Generating symphonic music requires simultaneously managing high-level structural form and dense, multi-track orchestration. Existing symbolic models often struggle with a "complexity-control imbalance", in which scaling bottlenecks limit long-term granular steerability. We present SymphonyGen, a 3D hierarchical framework for contemporary cinematic orchestration. SymphonyGen employs a cascading decoder architecture that decomposes the Bar, Track, and Event axes, improving computational efficiency and scalability over conventional 1D or 2D models. We introduce "short-score" conditioning via a beat-quantized multi-voice harmony skeleton, enabling outline control while preserving textural diversity. The model is further refined using Group Relative Policy Optimization (GRPO) with a cross-modal audio-perceptual reward, aligning symbolic output with modern acoustic expectations. Additionally, we implement a dissonance-averse sampling algorithm to suppress unintended tonal clashes during inference. Objective evaluations show that both reinforcement learning and dissonance-averse sampling effectively enhance harmonic cleanliness while maintaining melodic expression. Subjective evaluations demonstrate that SymphonyGen outperforms baselines in musicality and preference for orchestral music generation. Demo page: https://symphonygen.github.io/

Standard text-to-speech (TTS) evaluation measures intelligibility (WER, CER) and overall naturalness (MOS, UTMOS) but does not quantify accent. A synthesiser may score well on all four yet sound non-native on features that are phonemic in the target language. For Indic languages, these features include retroflex articulation, aspiration, vowel length, and the Tamil retroflex approximant (letter zha). We present PSP, the Phoneme Substitution Profile, an interpretable, per-phonological-dimension accent benchmark for Indic TTS. PSP decomposes accent into six complementary dimensions: retroflex collapse rate (RR), aspiration fidelity (AF), vowel-length fidelity (LF), Tamil-zha fidelity (ZF), Frechet Audio Distance (FAD), and prosodic signature divergence (PSD). The first four are measured via forced alignment plus native-speaker-centroid acoustic probes over Wav2Vec2-XLS-R layer-9 embeddings; the latter two are corpus-level distributional distances. In this v1 we benchmark four commercial and open-source systems (ElevenLabs v3, Cartesia Sonic-3, Sarvam Bulbul, Indic Parler-TTS) on Hindi, Telugu, and Tamil pilot sets, with a fifth system (Praxy Voice) included on all three languages, plus an R5->R6 case study on Telugu. Three findings: (i) retroflex collapse grows monotonically with phonological difficulty Hindi < Telugu < Tamil (~1%, ~40%, ~68%); (ii) PSP ordering diverges from WER ordering -- commercial WER-leaders do not uniformly lead on retroflex or prosodic fidelity; (iii) no single system is Pareto-optimal across all six dimensions. We release native reference centroids (500 clips per language), 1000-clip embeddings for FAD, 500-clip prosodic feature matrices for PSD, 300-utterance golden sets per language, scoring code under MIT, and centroids under CC-BY. Formal MOS-correlation is deferred to v2; v1 reports five internal-consistency signals plus a native-audio sanity check.

Commercial TTS systems produce near-native Indic audio, but the best open-source bases (Chatterbox, Indic Parler-TTS, IndicF5) trail them on measured phonological dimensions, and the most widely adopted multilingual base (Chatterbox, 23 languages) does not even tokenise Telugu or Tamil. We ask: what is the minimum intervention that brings such a non-Indic-native base to commercial-class output on Telugu, Tamil, and Hindi, without training a new acoustic decoder and without any commercial TTS training data? We combine three pieces: (1) BUPS, a Brahmic Unified Phoneme Space that deterministically romanises seven Indic scripts to ISO-15919 so Chatterbox's Latin tokeniser can process them; (2) a LoRA adapter on only the text-token predictor (Chatterbox's t3), trained on ~1,220h of licensed Indic audio with a Hindi-proxy language_id; (3) a voice-prompt recovery recipe -- an 8-11s same-language reference clip plus three sampling overrides (exaggeration 0.7, temperature 0.6, min_p 0.1; "Config B") -- that recovers commercial-class acoustic output with no acoustic-decoder training. On Hindi, the LoRA regresses accuracy and we instead use vanilla Chatterbox + Config B, giving a two-branch deployment. Evaluated on 10-utterance pilot sets with the companion PSP benchmark, Praxy Voice matches or slightly leads commercial baselines: 26.7% retroflex collapse on Telugu (vs Sarvam Bulbul 33.3%), 71% Tamil-zha collapse (vs commercial trio's 86%), 0.025 LLM-WER on Hindi (tied with Cartesia Sonic-3). For intra-sentential code-mix we add a third branch (IndicF5 + native-script transliteration) that drops code-mix LLM-WER from 0.80-0.85 to 0.14-0.27 across Hi/Te/Ta. We release R6 LoRA weights (Apache-2.0), inference code and router (MIT), and a Gradio demo.

To establish empathy with machines, it is essential to fully understand human emotional changes. However, research in multimodal emotion recognition often overlooks one problem: individual expressive traits vary significantly, which means that different people may express emotions differently. In our daily lives, we can see this. When communicating with different people, some express "happiness" through their facial expressions and words, while others may hide their happiness or express it through their actions. Both are expressions of 'happiness,' but such differences in emotional expression are still too difficult for machines to distinguish. Current emotion recognition remains at a 'static' level, using a single recognition model to identify all emotional styles. This "simplification" often affects the recognition results, especially in multi-turn dialogues. To address this problem, this paper introduces a novel Multi-Level Speaker Adaptive Network (ML-SAN), which, specifically, effectively addresses the challenge of speaker identity information confusion. ML-SAN does not simply assign a speaker's ID after recognition; instead, it employs a three-stage adaptive process: First, Input-level Calibration uses Feature-Level Linear Modulation (FiLM) to adjust the raw audio and visual features into a neutral space unrelated to the speaker. Then, Interaction-level Gating re-adjusts the trust level for each modality (e.g., voice or facial features) based on the speaker's identity information. Finally, Output-level Regularization maintains the consistency of speaker features in the latent space. Tests on the MELD and IEMOCAP datasets show that our model (ML-SAN) achieves better results, performs exceptionally well in handling challenging tail sentiment categories, and better addresses the diversity of speakers in real-world scenarios.

This performance presents a duet between two intelligent musical instruments, Sù (to trace back; to go upstream) and Agentier (playing on agentic clavier), and their human performers, connected through feedback loops. Rather than treating AI as a tool that responds predictably to input, both systems operate recursively, where past actions continuously influence future behaviour. The Sù operates in the audio space through latent representation. Its performer uses Make Noise 0-series synthesisers and MIDI controllers to work with a neural feedback synthesis system based on a RAVE model, with a latent feedback loop embedded within the model's internal structure. This allows the instrument to remember and reuse its own internal states, influencing ongoing sound generation through its recent sonic history. The Agentier functions in the control space. Its performer interacts with the system using a Roland S-1 synthesiser and Keith McMillen QuNeo touchpad, where control gestures are routed into a recurrent neural network that feeds back into the synthesis process. Through this feedback loop, the system actively shapes the evolution of control signals over time. Contrasting feedback in the audio and control domains, the performance explores shared agency, resistance, and negotiation between humans and intelligent musical systems. Musical phenomena are co-produced through the entangled states of interaction, rather than through pre-existing system configuration or fixed mappings.

Korean aegyo is a socially recognized childlike speaking style used predominantly in romantic interactions among adults. This study examined vowel space modification in aegyo by analyzing formant frequencies from twelve Seoul Korean speakers who produced identical scripts in aegyo and non-aegyo styles. Results show that aegyo speech features a significant increase in F1 values across vowels and selective fronting of front vowels, leading to vowel space expansion but mainly a shift to higher F1. These findings suggest that adult speakers stylize childlike speech by imitating the shorter vocal tract of children, mainly through global vowel lowering and partial fronting.

General audio foundation models have recently achieved remarkable progress, enabling strong performance across diverse tasks. However, state-of-the-art models remain extremely large, often with hundreds of millions of parameters, leading to high inference costs and limited deployability on edge devices. Knowledge distillation is a proven strategy for model compression, but prior work in audio has mostly focused on supervised settings, relying on class logits, intermediate features, or architecture-specific techniques. Such assumptions exclude models that output only embeddings, such as self-supervised or metric-learning models. We introduce S-SONDO (Self-Supervised KnOwledge DistillatioN for General AuDio FOundation Models), the first framework to distill general audio models using only their output embeddings. By avoiding the need for logits or layer-level alignment, S-SONDO is architecture-agnostic and broadly applicable to embedding-based teachers. We demonstrate its effectiveness by distilling two audio foundation models into three efficient students that are up to 61 times smaller while retaining up to 96% of teacher performance. We also provide practical insights on loss choice and clustering-based balanced data sampling. Code is available here: https://github.com/MedAliAdlouni/ssondo.

Despite recent progress in automatic speech recognition (ASR), elderly ASR (EASR) remains challenging due to limited training data and the distinct acoustic and linguistic characteristics of elderly speech. In this work, we address data scarcity in EASR through a data augmentation pipeline that combines large language model (LLM)-based transcript paraphrasing with text-to-speech (TTS) synthesis. Given an elderly speech dataset, the LLM first generates elderly-contextual paraphrases of the original transcripts, and the TTS model then synthesizes corresponding speech using elderly reference speakers. The resulting synthetic audio-text pairs are merged with the original data to fine-tune Whisper without architectural modification. We further analyze the effects of augmentation ratio and reference-speaker composition in low-resource EASR. Experiments on English and Korean elderly speech datasets from speakers aged 70 and above show that the proposed method consistently improves performance over conventional augmentation baselines, achieving up to a 58.2% reduction in word error rate (WER) compared with the Whisper baseline.

This paper presents StrADiff, a Structured Source-Wise Adaptive Diffusion Framework for unsupervised blind source separation under linear and nonlinear mixing. The framework treats each latent dimension as a source branch and assigns to it an individual adaptive reverse diffusion mechanism, so that latent sources are recovered directly from observed mixtures through a single end-to-end objective, without supervised source labels or separate post-processing. Source-wise generation, structural regularization, and observation-space reconstruction are optimized jointly during training. In this instantiation, a Gaussian process (GP) prior is used as one example of a source-wise structured prior to impose temporal organization on each recovered trajectory; the framework itself is not restricted to GP priors and can in principle incorporate other structured priors. Theoretical components clarify the induced pushforward source law, the sample-level role of the structured prior, the coupling between source recovery and prior adaptation, and a conditional weak recovery statement in an idealized linear low-noise regime. Experiments on linear and nonlinear mixtures show that StrADiff can recover meaningful latent source trajectories in an unsupervised manner, with particularly stable performance in the linear case and moderate degradation under nonlinear mixing. Beyond classical signal separation, a source branch may also be interpreted as an independent, disentangled, or otherwise interpretable explanatory factor under suitable structural assumptions, suggesting a broader route toward structured latent modeling and future identifiable nonlinear representation learning.

The Metaverse, a shared and spatially organized digital continuum, is transforming various industries, with music emerging as a leading use case. Live concerts, collaborative composition, and interactive experiences are driving the Musical Metaverse (MM), but the requirements of the underlying network and service infrastructures hinder its growth. These challenges underscore the need for a novel modeling and simulation paradigm tailored to the unique characteristics of MM sessions, along with specialized service provisioning strategies capable of capturing their interactive, heterogeneous, and multicast-oriented nature. To this end, we make a first attempt to formally model and analyze the problem of service provisioning for MM sessions in 5G/6G networks. We first formalize service and network graph models for the MM, using "live audience interaction in a virtual concert" as a reference scenario. We then present MuMeNet, a novel discrete-event network simulator specifically tailored to the requirements and the traffic dynamics of the MM. We showcase the effectiveness of MuMeNet by running a linear programming based orchestration policy on the reference scenario and providing performance analysis under realistic MM workloads.

Automatic Pitch Correction (APC) enhances vocal recordings by aligning pitch deviations with intended musical notes. However, existing APC systems either rely on reference pitches, which limits practical applicability, or employ simple pitch estimation algorithms that often fail to preserve expressiveness and naturalness. We propose BERT-APC, a reference-free APC framework that corrects pitch errors while maintaining the expressiveness and naturalness of vocal performances. In BERT-APC, a stationary pitch predictor first estimates the stationary pitch of each note from the detuned singing voice, where stationary pitch is the continuous pitch from the stable region of a note and approximates its perceived pitch. A context-aware note pitch predictor then infers the intended pitch sequence using a repurposed music language model that incorporates musical context. Finally, a note-level correction algorithm fixes pitch errors while preserving intentional deviations for emotional expression. We also introduce a learnable data augmentation strategy that improves robustness by simulating realistic detuning patterns. Compared to two recent singing voice transcription models, BERT-APC demonstrated superior target note pitch prediction, outperforming the second-best model, ROSVOT, by 10.49 percentage points on highly detuned samples in raw pitch accuracy. In the MOS test, BERT-APC achieved the highest quality rating of $4.32 \pm 0.15$, significantly higher than Auto-Tune ($3.22 \pm 0.18$) and Melodyne ($3.08 \pm 0.18$), while maintaining a comparable ability to preserve expressive nuances. To the best of our knowledge, this is the first APC model that leverages a music language model to achieve reference-free pitch correction with symbolic musical context. The corrected audio samples are available at https://joshua-1995.github.io/BERT-APC-Demo/.

Gesture-driven music generation is an emerging human-computer interaction paradigm for touch-free and expressive musical interaction. However, many existing approaches treat the task as isolated gesture classification or map gestures to symbolic outputs such as MIDI followed by a separate rendering stage, which limits temporal continuity and real-time responsiveness. This work presents Gesture2Music, a low-latency streaming framework for continuous gesture-driven music generation from live webcam feed. The system processes sequences of body and hand landmarks and uses a causal temporal convolutional network (TCN) to predict note-level musical control events, including pitch, octave, onset, sustain, amplitude, and activity state. Because available gesture-note datasets typically contain only isolated single-note recordings rather than continuous performance sequences, a synthetic stream generation strategy is introduced to construct continuous gesture streams by concatenating single-note clips and deriving heuristic temporal event labels. Temporal consistency and spectral proxy losses are further used to reduce prediction jitter and encourage audio-consistent outputs. During inference, predicted musical events are rendered into continuous music using predefined note samples with rhythmic quantization and scale-constrained filtering for improved musical stability. Experiments on a custom gesture-to-music dataset with 21 gesture-note classes spanning seven tones across three pitch levels demonstrate stable real-time performance, low inference latency of 30\,ms, and improved temporal continuity.

Human communication is multimodal, with speech and gestures tightly coupled, yet most computational methods for generating speech and gestures synthesize them sequentially, weakening synchrony and prosody alignment. We introduce Gelina, a unified framework that jointly synthesizes speech and co-speech gestures from text using interleaved token sequences in a discrete autoregressive backbone, with modality-specific decoders. Gelina supports multi-speaker and multi-style cloning and enables gesture-only synthesis from speech inputs. Subjective and objective evaluations demonstrate competitive speech quality and improved gesture generation over unimodal baselines.

Real-time audio-visual speech enhancement (AVSE) is a key enabler for immersive and interactive multimedia services, yet its performance is tightly constrained by network latency, uplink capacity, and computational delay. This paper presents the design, deployment, and evaluation of a complete cloud-edge-assisted AVSE system operating over a public 5G edge network. The system integrates CNN-based acoustic enhancement and OpenCV-based facial feature extraction with an LSTM fusion network to preserve temporal coherence, and is deployed on a Vodafone-compatible AWS Wavelength edge cloud. Through extensive stress testing, we analyze end-to-end performance under varying network load and adaptive multimedia profiles. Results show that compute placement at the network edge is critical for meeting real-time coherence constraints, and that uplink capacity is often the dominant bottleneck for interactive AVSE services. Only 5G and wired Ethernet consistently satisfied the required communication delay bound for uncompressed audio-video chunks, while aggressive compression reduced payload sizes by up to 80% with negligible perceptual degradation, enabling robust operation under constrained conditions. We further demonstrate a fundamental trade-off between processing latency and enhancement quality, where reduced model complexity lowers delay but degrades reconstruction performance in low-SNR scenarios. Our findings indicate that public 5G edge environments can sustain real-time, interactive AVSE workloads when network and compute resources are carefully orchestrated, although performance margins remain tighter than in dedicated infrastructures. The architectural insights derived from this study provide practical guidelines for the design of delay-sensitive multimedia and perceptual enhancement services on emerging 5G edge-cloud platforms.