arXiv_SD
-
Why Do You Say It Like That? A Phoneme-Level Framework for Explainable Speech Deepfake Detection
2026-07-09 15:23:45Anna Taylor, Michele Panariello, Massimiliano Todisco, Chiara Galdi, Nicholas Evans, Driss MatroufAbstract
As the accuracy of speech deepfake detection improves with the use of self-supervised representations such as wav2vec 2.0 and HuBERT, understanding why the speech is classified as bona fide or deepfake remains an open challenge. In pursuit of more trustworthy and interpretable artificial intelligence, we introduce a phoneme-level analysis framework that connects model predictions to measurable phonetic units. Our post-hoc explainability method is generally applicable to a variety of speech deepfake detection systems based on convolutional neural networks since it leverages Gradient-weighted Class Activation Mapping in conjunction with speech recognition to generate saliency maps aligned with phonemes and pauses. This pipeline reveals statistically significant attack- and speaker-dependent phonetic cues associated with spoofed speech in terms that humans can understand. Experiments using ASVspoof 5 show comparable detection performance to similar architectures while providing linguistic interpretations across speakers and spoofing conditions.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08586
PDF
-
Structural Bottlenecks on Frequency Representation in End-to-End Audio Models
Abstract
End-to-end neural audio models achieve high-fidelity compression and generation. We might read that performance as evidence they directly represent interpretable features such as pitch and timbre, but a model can produce plausible outputs without doing so. A model may encode these features in any reachable basis, but regardless of which, the features are well described as compositions of time-frequency-localized primitives. Whether state-of-the-art encoders preserve access to these primitives, and thus to compositions of them, remains unclear. Through theoretical analysis and controlled experiments, we show that several state-of-the-art strided convolutional encoders impose two structural bottlenecks, both predictable from architecture and signal structure, on access to these primitives: (1) they collapse primitives into alias equivalence classes, establishing a bound on representational capacity, and (2) they limit the frequency resolution available to learned filters, restricting separability. For well structured data, we find collapse rates of 31-35% and filter bandwidths 10-35x above the theoretical resolution bound, confirming that both bottlenecks arise under realistic signal conditions. We then introduce Gabor Latent Refactorization (GLRF), a lightweight post-hoc intervention that re-expresses encoder latents in a frequency-localized basis, reducing filter bandwidths from 10-35x to 1.5-3x of the theoretical resolution bound while preserving reconstruction fidelity and improving control over attributes like pitch. These results show that the encoders in question predictably degrade access to frequency-localized primitives, entangling the features that depend on them, and that a lightweight, retraining-free intervention can recover much of that access, improving steerability and interpretability.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08545
PDF
-
A Quantized Native Runtime for On-Device Semantic Audio Generation
Abstract
Semantic audio applications increasingly require controllable generation on commodity and embedded hardware rather than through framework-heavy datacenter stacks. We present \textit{aria}, a dependency-free native runtime that runs the complete text-to-music pipeline of Stable Audio~3 (SA3) on ordinary GPUs, CPU-only machines, and a Raspberry~Pi~5, with no Python or deep-learning framework underneath. Our main contribution is a study of quantization: running the model at lower numerical precision to fit tight memory budgets, saving memory in place rather than adding to it. Because the runtime owns every internal tensor, it also exposes activation steering, a low-cost way to steer what the model generates. We judge the quality cost with three independent measures of the output (prompt adherence, overall audio quality, taste preservation), each compared against the ordinary variation between random seeds. Eight-bit precision shows no measurable quality loss on any measure while sharply cutting memory, and it is the fastest mode on the GPU; four-bit adds a small, bounded cost but shrinks the footprint enough to run the $1.2$-billion-parameter model on an $8$\,GB Pi. Against the official implementation, aria matches or exceeds generation speed and starts about seven times faster. A case study of the steering interface generates music carrying taste associations (\emph{sonic seasoning}), with genuine but bounded control for a subset of attributes. These results make a compact, quantized runtime with built-in control a practical basis for on-device semantic audio in Internet-of-Sounds settings. The \textit{aria} runtime is released at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08526
PDF
-
Tonnetz-Driven Graph Wedgelet for Harmonic Complexity Reduction in Music Scores
2026-07-09 10:03:53Emmanuel Caronna, Elisa Francomano, Silvia LicciardiAbstract
Heterogeneous graph built on notes, lyric syllables, and accompaniment events is a natural representation of symbolic music score, providing a substrate for both philological analysis and computational tasks. Music features are therefore well-captured by graph geometry and its properties. This representation has proved effective for analytical tasks as cadence detection, voice separation, and stylistic classification. In the present work, the reduction of harmonic complexity of a music score on graph, by preserving task-relevant information, relation between notes, and graph structure is investigated. A compression scheme for the piano subgraph of vocal-pianistic scores, built on binary wedge partitioning trees, is proposed. The wedges are generated through a fully adaptive greedy algorithm that recursively minimizes the $L^2$-error within a six-dimensional Tonnetz embedding of musical notes. The partitioning process employs a splitting criterion based on harmonic distance, resulting in regions that accurately reflect the intrinsic harmonic relationships among notes. The reconstructed music scores obtained through piecewise-constant functions and the mean values of the notes inside each wedge are used as a new simplified scores human-readable and playable. Some experiments on a corpus of symbolic music scores of three different composers are performed to assess the proposed approach.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08806
PDF
-
MuScriptor: An Open Model for Multi-Instrument Music Transcription
2026-07-09 07:12:30Simon Rouard, Michael Krause, Axel Roebel, Carl-Johann Simon-Gabriel, Alexandre D\'efossezAbstract
Existing methods for automatic music transcription are often limited to single-instrument recordings or fail on complex, real music mixes. Although previous work utilizes synthetic training data, the resulting models generalize poorly, leading to largely unusable transcription output in realistic, multi-instrument settings. In this work, we analyze the effectiveness of synthetic data for pre-training while combining it with fine-tuning on real music audio and post-training using reinforcement learning. We further introduce conditioning on instrument presence to customize transcriptions. Finally, we release MuScriptor, an open-weight multi-instrument music transcription model that works on real-world music recordings from across a diverse range of musical genres.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08168
PDF
-
Dual-BEATs: Unlocking Zero-Shot Stereo Audio Perception in Audio Large Language Models via Dithering
2026-07-09 03:10:16Shuo-Chun Lin, Hen-Hsen HuangAbstract
Multimodal Large Language Models (LLMs) have remarkable semantic audio understanding, yet they remain "spatially agnostic" due to their reliance on mono-channel audio representations. Currently, spatial audio perception methods mainly focus on complex room simulations and custom-trained, geometry-aware stereo encoders, which limits their accessibility and generalizability. In this paper, we introduce the Dual-BEATs architecture, in which the left and right audio channels are routed independently through two identical semantic encoders as an alternative to specialized spatial modules. To circumvent the architectural bottleneck where internal normalization otherwise erases the inter-channel variance of stereo audio, we inject a static, uncorrelated dithering noise floor prior to encoding. This dithering intervention establishes a macro-variance floor that "smuggles" spatial geometry across the normalization layers. Evaluated on a ternary directional classification task (Left, Center, Right), we demonstrate that dithered models achieve exceptional spatial resolution--reaching up to 97.2% localization accuracy even on subtle 0.5 panning amplitudes--and demonstrates robust, zero-shot generalization to entirely unseen spatial configurations. Our results suggest that with the appropriate acoustic regularization, standard multimodal models are natively capable of generalized stereo audio understanding.
Abstract (translated)
URL
https://arxiv.org/abs/2607.08800
PDF
-
FlowTTS-GRPO: Online Reinforcement Learning with Multi-Objective Reward Optimization for Flow-Matching Based Text-to-Speech
2026-07-08 11:35:21Haoxu Wang, Biao Tian, Weiqin Li, Xiang Lv, Han Zhao, Xiangang LiAbstract
Existing Reinforcement Learning (RL) research for Text-to-Speech (TTS) focuses on large language models (LLMs), leaving Flow-Matching (FM) under-explored. We present FlowTTS-GRPO, an online RL framework for FM-based TTS. By converting ordinary differential equation (ODE) trajectories into stochastic differential equation (SDE) paths, our method enables direct fine-tuning of open-source FM models without auxiliary models. We show that a weighted reward combination converges faster than a probabilistic scheme, and identify three practical optimizations: omitting classifier-free guidance (CFG) during training accelerates convergence; synthesizing hard cases improves robustness; and applying RL to the FM component enhances audio-detail metrics. Experiments on CosyVoice 3.0 and F5-TTS demonstrate objective and subjective preference gains in speaker similarity and perceptual quality, with F5-TTS also improving intelligibility.
Abstract (translated)
URL
https://arxiv.org/abs/2606.23190
PDF
-
Rag Classification of Tagore Songs using Symbolic Music Notation and Novel Weighted Distance Measures
Abstract
Rabindra Sangeet, the body of songs written and composed by Rabindranath Tagore, occupies a distinctive position in Indian music by combining poetic expression with melodic ideas drawn from Hindustani rags, Bengali folk traditions, tappa, kırtan, Baul music, and Western tunes. Although many Tagore songs are associated with rag labels provided by Tagore himself or preserved in authoritative notational traditions, rag identification remains challenging because the songs often reflect creative freedom rather than strict adherence to classical rag grammar. This paper formulates rag identification in Rabindra Sangeet as a supervised classification problem using symbolic music-sheet notations from Swarabitan. Since large-scale annotated audio or music datasets for Rabindra Sangeet are not readily available, this study constructs a rag-labelled symbolic dataset from notated Tagore songs. The work investigates Euclidean distance and cosine similarity for rag classification and introduces a weighted Euclidean distance measure that assigns greater importance to notes belonging to characteristic rag sequences such as arohana and avarohana. Applied within a k-nearest-neighbour framework, the proposed measure improves rag classification by better capturing rag-specific melodic identity.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07241
PDF
-
EscFOA: Enhancing Spatial Learning for Visually Impaired Learners via Generative Spatial Audio in 360-Degree Educational Environments
2026-07-08 05:17:26Ziyu Luo, Xiaowei Dai, Siying Zhu, Xiaoming ChenAbstract
Immersive 360-degree educational environments often lack accessible spatial structure, limiting visually impaired learners' ability to orient, explore, and construct mental representations. This paper proposes EscFOA, a geometry-aware spatial audio generation framework designed as an \emph{acoustic scaffolding} to support spatial cognition. By integrating 3D Gaussian Splatting (3DGS) with conditional diffusion models, EscFOA reconstructs scene geometry from 360-degree videos to synthesize high-fidelity spatial audio consistent with the environmental structure. Explicitly targeting learning outcomes like independent spatial orientation and reduced cognitive load, EscFOA significantly outperforms conventional monaural and stereo audio in supporting spatial learning behaviors among blindfolded sighted participants (simulating visually impaired learners). These findings demonstrate that geometry-consistent generative audio can effectively enable inclusive access to complex spatial learning materials.
Abstract (translated)
URL
https://arxiv.org/abs/2607.07015
PDF
-
MMGenre: Benchmarking Singing Voice Synthesis across Multiple Musical Genres
Abstract
Singing voice synthesis (SVS) has progressed rapidly, yet its ability to generalize across diverse musical genres remains underexplored. Existing benchmarks are heavily biased toward pop music, limiting systematic analysis of genre-dependent behavior. We introduce MMGenre, a benchmark for multi-genre SVS diagnosis, supported by an automatic pipeline for constructing genre-aligned music scores. MMGenre spans 10 major genres and 26 subgenres, enabling comprehensive analysis of genre-aware synthesis. Extensive evaluation of representative SVS models reveals limited genre discrimination: synthesized vocals across genres exhibit highly similar acoustic characteristics and weak separability. While zero-shot genre adaptation yields only marginal improvements, lightweight genre-specific continued training leads to substantial gains. MMGenre provides a standardized framework for multi-genre SVS evaluation and exposes critical challenges in achieving genre-aware singing voice synthesis.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06986
PDF
-
Compress the Cache, Not the Speech Embedding: KV Compression for Efficient Speech LLMs
2026-07-07 21:45:04Ke-Han Lu, Keqi Deng, Ruchao Fan, Rui Zhao, Jinyu LiAbstract
Speech large language models (Speech LLMs) typically encode speech into sequences far longer than text, creating a major efficiency bottleneck during autoregressive decoding. A common remedy is to compress the speech sequence at the adapter level to remove temporal redundancy before it enters the LLM; however, such early downsampling risks discarding fine-grained information that cannot be recovered. We propose SpeechKV, which applies a learned pooling to the KV cache of speech tokens inside the LLM. This design allows the LLM to fuse speech and text internally while directly accelerating decoding. Trained on 71K hours of speech data, SpeechKV compresses the speech to approximately text-level granularity yet maintains performance on par with or even slightly better than the uncompressed baseline, with relative gains of 6.6% on out-of-domain entity recognition and 2.3% on OpenASR, while delivering at least 1.49 times decoding speedup that scales with audio length.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06827
PDF
-
Precise Video-to-Audio Generation with Cross-Modal Alignment in Latent Space
2026-07-07 15:34:43Thanh V. T. Tran, Ngoc-Son Nguyen, Luong Tran, Long-Khanh Pham, Paarth Neekhara, Shezheen Hussain, Van NguyenAbstract
Video-to-audio (V2A) generation aims to synthesize realistic audio that is both semantically consistent with and temporally synchronized to a silent video. Despite recent progress, many methods still rely on multi-stage training, resulting in high computational costs and long runtimes, or transform visual input into text to leverage pretrained text-to-audio models, sacrificing fine-grained temporal cues. To overcome these limitations, we propose Flowley, an end-to-end, single-stage training architecture that produces soundtracks by combining visual features with textual prompts. Crucially, we introduce Progressive Soft-masked Cross-Attention, which embeds audio-visual synchronization directly within its attention mechanism, adding zero additional computational cost compared to standard attention layers. We further observe that existing V2A benchmarks lack sound-oriented descriptive captions, which can potentially degrade the quality of the synthesized audio. To remedy this, we propose SoundCap, a plug-and-play pipeline for creating detailed, sound-aware captions that guide the model. Remarkably, without integrating any pretrained audio-visual alignment modules, Flowley achieves state-of-the-art performance on VGGSound across multiple metrics. Moreover, by incorporating SoundCap, we further exceed the performance of the strongest existing close-sourced methods in terms of audio quality in the zero-shot setting.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06405
PDF
-
InsideSSL: Understanding Self-Supervised Speech Representations using a Model-Centric Perspective
2026-07-07 15:25:47Samir Sadok, Xavier Alameda-PinedaAbstract
Self-supervised learning (SSL) models, such as Wav2Vec2, HuBERT, and WavLM, have become foundational across a wide range of speech and audio tasks. Despite their success, understanding their internal layer-wise dynamics remains an ongoing challenge. To address this, we propose a two-part model-centric framework called InsideSSL. First, we establish a task-agnostic analysis from three intrinsic per-layer perspectives: compression (entropy), geometry (curvature), and robustness to perturbations. We show that varying training objectives induce distinct regimes of acoustic compression and manifold unfolding. Second, we introduce the cross-layer Generative Compatibility Matrix (GCM) to evaluate functional transferability, exposing stable phonetic cores, identity volatility, and deep-layer semantic pruning. In addition to these evaluations, linear probing connects the model-centric perspective to downstream tasks, demonstrating how layer topology dictates phoneme, pitch, and speaker encoding.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06392
PDF
-
ForestIR: Physics-Informed Forest Sound Simulation for Array-Based Bioacoustic Remote Sensing
2026-07-07 14:04:25Xin Shen, Jennifer N. Kampe, Changwoo J. Lee, Braden Scherting, Panu Somervuo, Ari Lehti\"o, Sandro von Brandenburg, Ossi Nokelainen, Otso Ovaskainen, David B. DunsonAbstract
Microphone array-based passive acoustic monitoring is increasingly used for biodiversity sensing in forests. However, design and evaluation of array systems and configurations remains difficult since field recordings are costly, difficult to reproduce, and provide limited control over forest and atmospheric conditions. We present ForestIR, a physics-informed and reproducible simulation framework that links forest and environmental conditions to microphone-array recordings for bioacoustic remote sensing. Through a more realistic sound propagation method and a systematic control over array design and environmental factors, ForestIR provides a practical simulation framework for optimizing array-based monitoring systems, especially for sound source localization purposes. ForestIR generates source-microphone impulse responses (IRs) under user-controlled forest and atmospheric conditions, and renders synthetic array recordings by convolving test signals with controlled background noise. We evaluate and demonstrate realistic features of ForestIR through experiments based on localization sensitivity to forest layout and atmospheric conditions, and also comparison between simulated IRs with sine-sweep IR measurements from a field experiment. ForestIR provides a practical way to test how forest and ground conditions, atmospheric state, and array geometry affect bioacoustic localization, and can support microphone-array design, robustness testing, and synthetic-data generation for passive acoustic monitoring.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06299
PDF
-
Learning-based Physics-Constrained Neural Kernel for Sound Field Estimation With Source-Position-Dependent Directional Weighting
Abstract
A learning-based physics-constrained neural kernel for sound field estimation is proposed. Sound field estimation aims to estimate the spatial distribution of an acoustic field from a discrete set of microphone measurements, which have a wide range of applications. Among existing sound field estimation methods, kernel-regression-based methods offer a flexible and principled framework for incorporating physical constraints and allow inference through linear operation. It is also possible to adapt the kernel function to the target acoustic environment by representing the directional weighting function as an implicit neural representation (INR) and optimizing hyperparameters using measurements. However, the kernel function is generally optimized for single snapshot measurements of the microphones, which can lead to strong overfitting and poor generalization. We propose a source-position-dependent INR for the directional weighting function, enabling the kernel function to capture common directional patterns and to generalize to unseen source positions in the target acoustic environment. Experimental results indicate that our proposed method outperforms the snapshot-based method by estimating a directional weighting function that matches the directivity of the target sound field.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06274
PDF
-
Flow Matching-Based Speech Source Separation with Best-of-N Biometric Sampling
2026-07-07 10:00:23Anastasia Zorkina, Alexandr Anikin, Nikita Khmelev, Anastasiya Korenevskaya, Sergey Novoselov, Vladimir Volokhov, Maxim Korenevsky, Yuriy MatveevAbstract
Single-channel speech separation remains challenging for real-world deployment due to source permutation ambiguity, sampling variability of generative models, and the difficulty of processing long recordings with chunk-wise inference. We address these issues with a conditional flow-matching-based method that produces an ordered two-source output conditioned on the mixture. A frozen speaker encoder defines the source order during training and is reused at inference for biometric best-of-$N$ candidate selection and chunk-level channel alignment. We evaluate separation quality on Libri2Mix benchmark using SI-SDR, PESQ, and ESTOI, and measure downstream impact using cpWER for automatic speech recognition and EER for speaker verification. The results show that the proposed Transformer U-Net variant is competitive with strong baselines in objective separation metrics and achieves the lowest downstream automatic speech recognition and speaker verification error rates in all evaluated settings.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06088
PDF
-
Determinantal point process sampling for bioacoustic active learning
Abstract
Eco-acoustic monitoring generates vast volumes of audio data, making active learning a promising approach for reducing annotation effort while efficiently training reliable biodiversity classifiers. This report presents CARE-DPP, a batch active-learning acquisition method submitted to BioDCASE Active Learning for Bioacoustics 2026 challenge. The method combines class-balanced predictive uncertainty with embedding-space novelty, while a determinantal point process (DPP) objective selects a high-quality and non-redundant acquisition batch. The uncertainty-novelty balance is annealed over the annotation budget: early cycles emphasize geometric coverage, whereas later cycles increasingly exploit classifier uncertainty. To mitigate unreliable early scores, the DPP candidate pool mixes top-quality candidates with a decreasing proportion of random exploration. An adaptive acquisition schedule uses smaller batches early and larger batches later. Evaluated over five repeats on the BirdSet HSN, POW and UHH subsets and on ATBFL, CARE-DPP obtains a mean development AULC of 0.50 for macro mAP, compared with 0.46 for the official CoreSet baseline. Ablations identify DPP batch diversification and the adaptive acquisition schedule as the largest contributors.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06063
PDF
-
Fr'echet Distance Loss on Speech Representations for Text-to-Speech Synthesis
2026-07-07 09:07:30Ho-Lam Chung, Kuan-Po Huang, Bo-Ru Lu, Hung-yi LeeAbstract
Few-step diffusion and flow-matching text-to-speech (TTS) models are usually trained with local objectives, such as conditional flow matching, reconstruction, and stop prediction. These losses provide stable optimization, but they never ask whether sampled speech follows the distribution of high-quality speech. We propose Speech Representation Fr'echet Distance loss (SR-FD), a training-time distributional regularizer for tokenizer-free flow-matching autoregressive TTS. During fine-tuning, the model synthesizes speech with the same few-step sampler used at deployment, and SR-FD matches the mean and covariance of frozen Whisper and CTC features of this speech to reference statistics computed offline from three complementary content targets. The loss requires no discriminator and no inference-time computation. On Seed-TTS English, four-step SR-FD fine-tuning reduces WER from the original four-step VoxCPM2 baseline's 2.2279% to 1.4147%, a 36.5% relative reduction, and also surpasses the original ten-step baseline at 1.7366%; both gains are significant under an utterance-level paired bootstrap. Speaker similarity and objective quality proxies are preserved at the ten-step level, and an error analysis shows the gain comes from content substitutions across all prompt lengths. SR-FD is thus an intelligibility-improving distributional regularizer for few-step TTS.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06027
PDF
-
Music I Care About: Automated Multimodal Benchmarking of LLM Music Perception Skills on Any Music
2026-07-07 08:57:34Tom\'a\v{s} Sourada, Katia Vendrame, Jan Haji\v{c} jrAbstract
Music represents a cornerstone of human culture, existing digitally across diverse modalities, including audio, symbolic encodings (e.g., MIDI, MusicXML), and sheet music. Despite the advancement of Multimodal Large Language Models (MLLMs), current music benchmarks face three major limitations. First, large static benchmarks are resource-intensive to evaluate, and it remains unclear how their results transfer to diverse kinds of music beyond those included in the benchmark. Second, benchmarks claiming to measure "music understanding" often fail to require music perception. Third, they do not support systematic performance comparisons across musical modalities. To overcome these issues, we introduce the Music I Care About Meta-Benchmark (MusICA-MetaBench), a framework that automatically derives on-demand benchmarks directly from user-provided data. By leveraging structured symbolic representations (e.g., MusicXML) and our pre-defined question templates, we build multiple-choice question-answer pairs that probe music perception competencies, aligned with music pedagogy, across audio, music notation images, and symbolic files. We demonstrate our framework with the ChoraleBricks dataset, and experimentally determine benchmark sizes that ensure statistically reliable model comparisons for this setup. By comparing against text-only and white-noise baselines, we show our questions do measure music perception. Ultimately, MusICA-MetaBench represents a significant advancement in the cross-modal assessment of music perception for MLLMs. By proposing a dataset-specific benchmarking paradigm, it enables efficient on-demand evaluation of music perception capabilities.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06015
PDF
-
Escaping the Procrustean Bed: Groupwise Orthogonal Connectors for Audio-Language Models
2026-07-07 08:57:00Ho-Lam Chung, Ke-Han Lu, Yi-Cheng Lin, Guan-Ting Lin, Yiming Chen, Hung-yi LeeAbstract
Audio-language models compress a speech encoder's output through a Querying Transformer (Q-Former) connector before feeding it to a large language model. We identify two failures in this compression. The connector's output vectors collapse to a single direction, and different speakers produce nearly indistinguishable outputs, with paralinguistic cues such as speaker identity, gender, and prosody lost along the way. Our method, ORCA, reverses this collapse by splitting the queries into groups whose outputs are constrained to point in different directions. On SAKURA multi-hop reasoning, ORCA gains 26.4 points over an identically trained 4B baseline, reaching 75.2% (vs. 49.0% for the 8B Audio Flamingo-3). At the connector level, the same change cuts query redundancy by 12x and raises cross-speaker variance by 75x.
Abstract (translated)
URL
https://arxiv.org/abs/2607.06014
PDF
