Although automatic detection of cognitive impairment from conversational speech is a promising way, it becomes less accurate, as large real-world datasets often contain unreliable diagnostic labels due to self-referral and limited clinical verification, and here we investigated whether neuropsychologically calibrated diagnostic relabeling improves it. We analysed 138 hours of conversational speech from 720 participants, recalibrating unreliable diagnostic labels using attention, executive function, and memory performance via external clinical thresholds, Gaussian, and normative z-score calibration. Transformer-based classifiers were trained using both original and recalibrated labels and evaluated internally on 150 clinically verified participants and externally on Pitt, ADReSS and PROCESS-2 datasets. Memory- and MoCA-based z-score calibration produced the greatest improvements. Using our internal dataset, $F_1$ increased from $0.70$ to $0.83$ ($+13\%$) and from $0.54$ to $0.66$ ($+12\%$) for binary and multiclass classification, respectively. For external validation, the increases are $+4\%$, $+5\%$ and $+9\%$ on Pitt, ADReSS and PROCESS-2 corpus. Neuropsychologically calibrated relabeling substantially improves speech-based cognitive assessment, highlighting label quality, rather than model architecture, as a critical bottleneck in clinical AI for dementia screening.
https://arxiv.org/abs/2610.28288
Reinforcement learning can improve text-to-speech synthesis through feedback on generated speech. However, sentence-level rewards do not identify local errors, while signed regression updates can lead to excessive fitting or repulsion of sampled outputs. We propose LoCA (Local Credit Assignment) for autoregressive flow-matching text-to-speech. LoCA uses ASR-based text alignment to locate content errors in generated speech and redistribute sentence-level advantages across acoustic frames. To improve training stability, we introduce reference-relative positive and negative thresholds that limit continued fitting and repulsion. Experiments on VoxCPM2 show that LoCA reduces sample-weighted content error on a multilingual CV3 evaluation subset by 20.5% and 14.3% relative to the pretrained backbone and GROW, respectively, while improving speaker similarity. LoCA also improves content accuracy on the Seed-TTS-Eval hard subset. Ablation results support the contributions of local credit and threshold-based regression control.
https://arxiv.org/abs/2610.28035
Real-world audio typically contains concurrent sound events, yet standard self-supervised methods treat each clip as an isolated training unit without explicitly modeling mixture--constituent relationships. We propose Mixture-to-Constituent Latent Alignment (MiCLA), a two-stage self-supervised pre-training framework that uses synthetic log-mel mixing to align mixture representations with their constituent clips. Stage 1 structures single-clip representations via masked self-distillation and contrastive regularization. Stage 2 introduces bidirectional alignment: for a synthetic mixture, MiCLA suppresses the component aligned with one constituent and aligns the residual with the other, using projective residualization on patch tokens and sample-adaptive soft residualization on the global class token. In downstream evaluations, MiCLA achieves 43.2% mean average precision (mAP) on AS-20K and 50.3% mAP on AS-2M under fine-tuning, and achieves the highest mAP among evaluated models on both AudioSet settings under frozen Convex Gated Probing (CGP). Furthermore, a Joint Semantic Retrieval diagnostic shows that MiCLA improves simultaneous retrieval of both constituent classes over target-averaging baselines. These results indicate that directional mixture-to-constituent alignment provides an effective pre-training signal for polyphonic audio representation learning.
https://arxiv.org/abs/2610.27491
Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.
https://arxiv.org/abs/2610.27210
Scene-adaptive immersive audio rendering requires rapid estimation of acoustic parameters that reflect the surrounding environment. For repeated runtime scene updates, performing a detailed geometric-acoustic simulation for every update can be computationally costly. We present Point2Acoustic, a learning-based framework for direct estimation of six-band reverberation time ($RT_{60}$) and reverberant-to-direct ratio (RDR) from 3D scene geometry and acoustic material properties. Each scene is represented by a surface-sampled point cloud with spatial coordinates and frequency-dependent absorption and scattering coefficients, and encoded using PointNet or PointNet++. We train and evaluate the framework on a synthetic furnished-room dataset with varied geometry and acoustic materials, using geometric-acoustic simulation to obtain reference targets. Point2Acoustic achieves accurate multi-band prediction with millisecond-scale inference, and ablation results indicate that explicit global room descriptors and semantic surface labels provide no consistent benefit. The proposed formulation supports efficient scene-adaptive acoustic rendering in extended reality (XR).
https://arxiv.org/abs/2610.27023
Sanskrit verse recitation is governed by $\textit{Chandas}$, a strict prosodic framework defined by rhythmic $\textit{Laghu}$ (light) and $\textit{Guru}$ (heavy) syllables. Because existing Sanskrit speech datasets prioritize ASR or conversational TTS, they lack the metrical fidelity required for authentic verse synthesis. We introduce $\textbf{\textit{ShlokaDhara}}$, a 43-hour, multi-speaker Sanskrit speech corpus capturing diverse meters, recitation styles, and melodic variations. We specifically focus on $\textit{Text-to-Recite}$, a synthesis task explicitly requiring both prosodic accuracy and metrical correctness. Benchmarking state-of-the-art TTS architectures via an automated UTMOS (neural MOS) predictor demonstrates that XTTS-v2 achieves the highest perceptual quality. We further validate the XTTS-v2 generated samples using two paradigms: subjective assessments by 16 Sanskrit experts and novel empirical metrics, specifically, syllable-level $\textit{Laghu-Guru}$ duration analysis to objectively quantify structural metrical awareness. Ultimately, this work establishes a robust dataset and evaluation benchmark for prosody-aware computational Sanskrit recitation. Generated Samples: this https URL
https://arxiv.org/abs/2610.26929
Interactive models increasingly hand complex reasoning and tool use to a slow brain in the background, leaving a realtime front model free to sustain full-duplex dialogue. Yet the user may refine, cancel or replace a request while its task runs, so a plan or result that made sense a moment ago goes stale. Each new utterance must therefore be read against the work in flight, but the front model follows only the conversation and the back end sees only its own task. Coordination deserves a brain of its own, one that sees both. We introduce Jarvis, a coordination harness where models judge and code enforces. Its typed decision model, JEV, reads each turn against the live tasks and picks among operations built from a versioned task record. Code checks every choice against the current state before applying it, so no decision can act on a task version the user has replaced. Jarvis also rechecks plans before they execute and results before they are spoken, and counts an answer as delivered only once the user has heard it in full. We build two instances of this design. Jarvis-Audio runs unchanged on three full-duplex speech models. Jarvis-Omni adds vision and rechecks the evidence behind an answer before and during playback, since a scene can change while nobody speaks. With the front model and back end held fixed, replacing Qwen Audio Agent with Jarvis raises tau-Voice task success more than sixfold, from 29 to 192 of 278 tasks, and the gains hold on every benchmark and front model. Against published systems, Jarvis-Audio reaches 69% Pass1 on Full-Duplex-Bench v3 with an open 9B front model, versus 60% for GPT-Realtime. These results suggest that how the two brains are coordinated matters as much as how strong they are.
https://arxiv.org/abs/2610.23804
Self-supervised speech models such as WavLM achieve strong spoofing detection performance, yet which layers encode the most spoof-relevant information and what acoustic properties they capture remain opaque. We present a layer-wise interpretability study of WavLM using the ASVspoof 5 database, combining three complementary analyses: (i) per-layer linear probing with attack-level evaluation across speech synthesis, voice conversion, and adversarial attacks; (ii) a canonical correlation analysis (CCA) linking layer representations to spoofing-relevant acoustic features; (iii) an exhaustive grid search over layer subsets showing that use of five layers outperforms full thirteen-layer pooling. CCA analysis reveals that, while most acoustic features peak in alignment with mid-layers and then decay, voice quality measures jitter, shimmer, and the harmonics-to-noise ratio show the opposite trend, with correlation increasing toward later layers. Embedding visualisations reveal that deeper layers capture a continuous acoustic similarity among attacks rather than discrete categories, suggesting these layers encode fundamental synthesis artifacts rather than attack-specific signatures.
https://arxiv.org/abs/2610.23526
The recurrent neural network transducer (RNN-T) remains a practical choice for streaming speech recognition, but training requires large intermediate tensors. We present Flash RNN-T, a joint and loss implementation combining specialized kernels with optional tiled recomputation over the full alignment space and vocabulary. Tiled and untiled Flash form the entire measured throughput-memory frontier on medium and long workloads. On H100, untiled Flash achieves 24-39% higher throughput with 26-56% less memory than the fastest baseline on these workloads wherever it fits. Tiling reduces memory further through recomputation. On medium, tiled Flash uses ninefold less memory than that baseline at 16% lower throughput when both use one backward call per batch. On long, it fits where all tested alternatives with that backward schedule run out of memory with 80 GB. These comparisons use each implementation's lowest-memory configuration reaching at least 95% of its peak throughput.
https://arxiv.org/abs/2610.23409
Dysarthric audio-visual speech recognition (DAVSR) enhances performance by integrating complementary visual cues. However, existing research is largely confined to word-level tasks with limited vocabularies, which severely restricts real-world communication for dysarthric speakers. Furthermore, performance remains heavily dependent on supervised fine-tuning with speaker-specific data; significant variations in age, etiology, and severity cause traditional methods to struggle with unseen speakers. To address these domain shifts and the scarcity of continuous multimodal resources, we first introduce a "source speech search-segmentation" pipeline to construct a structured sentence-level corpus from raw CDSD videos. We then propose the use of unsupervised Test-Time Adaptation (TTA) for sentence-level DAVSR using Sequential-level Generalized Entropy Minimization (SGEM). Our evaluations across the state-of-the-art AV-SUPERB and ViSpeR frameworks demonstrate that TTA consistently alleviates Character Error Rate (CER) degradation for unseen pathological speakers. Paired t-tests verify that these improvements are statistically significant (p < 0.05), highlighting TTA as a promising paradigm for advancing speaker-adaptive assistive speech interfaces.
https://arxiv.org/abs/2610.23117
Mispronunciation detection and diagnosis (MDD) provides phone-level feedback to language learners, but extending it to new languages typically requires collecting and annotating learner speech for each target language. This repeated annotation effort is costly and limits the language coverage of pronunciation feedback. We study whether MDD capabilities learned from annotated learner speech can be applied to languages unseen during MDD training when canonical pronunciation references are available. Our multilingual LLM-based model, X-MDD, uses canonical speech and phones as references to predict the phone sequence actually produced by a learner. It incorporates in-context demonstrations for inference-time adaptation and is further optimized with group relative policy optimization (GRPO) using detection F1 and phone error rate. We further analyze how canonical inputs affect prediction and find that (1) canonical speech and phones jointly contribute to actual-phone prediction, and (2) conflicting speech evidence reduces the tendency to copy canonical phones. Trained only on Mandarin and English, X-MDD achieves a detection F1 of 38.07 on Arabic Qur'anic recitation, compared with 44.14 for a supervised challenge baseline trained on Arabic. It also achieves F1 scores of 78.68 and 58.78 on Mandarin and English, respectively, demonstrating effective multilingual MDD within a single model.
https://arxiv.org/abs/2610.23079
Target speaker extraction (TSE) isolates a desired speaker from a multi-speaker mixture using cues that identify the target. Prior work has explored acoustic and textual cues, such as enrollment audio and different types of speaker descriptors, but typically relies on a single cue family. Yet no single cue is both broadly available and reliably discriminative. Moreover, most TSE systems adopt mask-prediction models, which are prone to audible artifacts and generalize poorly across acoustic conditions. To address these limitations, we propose PromptSep-Speaker, a generative TSE model based on latent diffusion that unifies four interchangeable cues: speaker temporal order, transcript, enrollment audio, and speaker caption. The generative formulation delivers higher separation quality and multi-modal prompting improves applicability across scenarios. To fully exploit each cue, we further introduce a frozen WavLM representation with an auxiliary alignment loss that supervises query-mixture cross-attention without extra labeling. Across four evaluation sets and a listening study, PromptSep-Speaker achieves state-of-the-art performance over TSE and universal source separation baselines, improving separation accuracy and quality across cue modalities.
https://arxiv.org/abs/2610.23066
Dynamic music emotion recognition usually estimates concurrent or lag-aligned labels from audio that has already been observed. We instead forecast the change in consensus arousal when the upcoming music is scheduled but all later affect labels remain hidden. Evaluation holds out complete songs, encodes each eligible 4-s chunk independently, predicts arousal innovation, and compares the correct future with displaced segments from the same song. On PMEmo and a 1,494-song DEAM evaluation pool excluded from engineering development, a complete past/future MuQ representation raises concordance correlation coefficient (CCC) by 0.389--0.507 over recent affect and remains 0.268--0.382 above a calibrated nonlinear state model. A future-only trajectory contains no past or current audio and still exceeds that state model by 0.150--0.249. The earliest legal future chunk, which ends 5.5--11 s before the target, also remains informative. Correct placement exceeds an overlapping 5-s shift by 0.100--0.209 and the acoustic-nearest eligible same-song segment by 0.078--0.105, with positive song-cluster lower bounds. These results show that both scheduled content and its within-song position carry forecast information under the stated observation rules. The study provides a leakage-controlled reference for known-future music forecasting without treating predictive substitutions as causal interventions.
https://arxiv.org/abs/2610.23063
Evaluating musical perception in multimodal language models requires separating audio evidence from answers recoverable through metadata or external tools. We present the Music Understanding Benchmark (MUB), a locally assembled collection of 500 identity-masked recordings with existing annotations for tempo, key, time signature, and structural intervals. The collection covers all 24 major/minor key combinations, tempos of 47--180 beats per minute, and 4,676 annotated intervals. Original titles, artists, and platform identifiers are excluded from evaluation inputs. Relative to recent music question-answering and temporal-grounding work, MUB emphasizes scalar tempo prediction and full-recording interval outputs, together with a documented metadata-masking protocol. Saved baseline predictions yield 53.6\% tempo accuracy for a simple DSP estimator and 98.4\% time-signature accuracy for a constant 4/4 predictor. We identify important limitations, including an asymmetric legacy key metric, unverified independent annotation agreement, extreme metre imbalance, and unresolved recording-level rights clearance. No contemporary multimodal language model comparison has yet been completed. MUB is therefore an evaluation resource and protocol under validation, rather than a completed demonstration of model capabilities.
https://arxiv.org/abs/2610.22989
Spectral losses for differentiable digital signal processing (DDSP) commonly compress magnitudes, for example by using log-magnitude terms. Such compression mimics human perception, but whether it helps the optimization is less clear, and is usually found only by comparing trained models. In this work, we show that whether compression helps depends on the system's informative region: the region of its dynamic range that carries the information pointing toward the target. To indicate where this region lies without having to train, we introduce concordance, which measures how well a loss orders candidate parameter settings by their distance from the target. We then compare uncompressed, compressed, and combined losses by fitting a modal plate reverb and a subtractive synthesizer model to target sounds. A third system, selected for its low-energy informative region, tests the relationship in the opposite direction. In every system, the loss that matches the informative region performs best: the uncompressed one where the region lies at high energies, and the compressed one where it lies low. Surprisingly, the common combination of both log-magnitude and linear-magnitude terms performs best in none of the systems. Code and audio examples are available at this https URL
https://arxiv.org/abs/2610.22962
Text-to-speech models built on score matching and flow matching produce natural speech, but only by calling the decoder many times, which makes synthesis slow. Existing ways to cut the number of calls either distill from a multi-step teacher or add a stage on top of a pretrained model, leaving single-step TTS without a pretrained TTS model unexplored. We question whether pretraining a TTS model is necessary, and propose DriftTTS, which adopts generative drifting, an objective that moves iterative sampling from inference into training. Drifting assumes samples of fixed shape, whereas speech has variable length; we address this issue with segment drifting. On LJSpeech, DriftTTS matches the multi-step quality of Matcha-TTS and RapFlow-TTS in a single decoder call, while being faster. To our knowledge, this is the first application of generative drifting to text-to-speech.
https://arxiv.org/abs/2610.22960
Automated respiratory sound classification (RSC) aims to identify normal and adventitious respiratory sounds, such as crackles and wheezes. Most existing methods formulate RSC as a flat four-class classification problem over norma, crackle, wheeze and both. However, unlike conventional classification tasks with semantically independent and mutually exclusive categories, the commonly used four-class RSC labels have a compositional structure and can be represented by two clinically meaningful sound attributes: crackle presence and wheeze presence. Existing flat classifiers do not explicitly exploit this structure, which may increase the learning burden by treating the four labels as independent targets. Meanwhile, crackles and wheezes exhibit distinct temporal characteristics, with sparse transient crackle events and sustained continuous wheeze patterns, making a single temporal aggregation strategy less suitable for modeling different adventitious sound evidence. To address these limitations, we propose an Acoustic-Aware Asymmetric Framework (AAAF) for RSC. AAAF reformulates four-class recognition as two sound-attribute detection tasks and recovers the final category through logical composition during inference. Built upon a pretrained BEATs encoder, AAAF employs branch-specific multi-layer fusion and asymmetric temporal aggregation to capture localized crackle evidence and sustained wheeze patterns. Auxiliary normal--abnormal supervision and a batch-adaptive dynamic loss are further introduced to improve global discrimination and robustness under imbalanced labels. Experiments on ICBHI and SPRSound demonstrate that AAAF improves overall RSC performance and generalizes across datasets, achieving a state-of-the-art four-class Score of 65.30% on ICBHI.
https://arxiv.org/abs/2610.22854
Respiratory sound classification distinguishes four categories: Normal, Crackle, Wheeze, and Both (crackle and wheeze). Existing knowledge distillation transfers four-class predictions without explicitly exploiting shared respiratory events, potentially limiting the transfer of event-specific knowledge across individual abnormal classes and the Both class. We propose PAC-KD, a pathology-aware compositional knowledge distillation framework that learns crackle and wheeze presence through two dedicated specialists. Their complementary knowledge is transferred to a unified four-class student, enabling it to effectively capture pathology-specific attributes while preserving class-level discrimination. Furthermore, stable conflict calibration adaptively regulates specialist guidance according to its consistency with classification supervision, mitigating potential supervision conflicts. Experiments on ICBHI and SPRSound achieve Scores of 65.28\% and 89.54\%, respectively, outperforming strong knowledge distillation baselines on both datasets.
https://arxiv.org/abs/2610.22826
We introduce TRACE (Trustworthy Reasoning from Acoustic Cues and Evidence), a benchmark for diagnosing the reliability of large audio-language models (LALMs) by revealing when, and under what acoustic conditions these models behave untrustworthily. Existing benchmarks primarily measure whether a model produces the correct answer, but correctness alone cannot determine whether the answer is grounded in acoustic evidence or instead reflects language and contextual priors. TRACE evaluates model behavior along three complementary dimensions: (1) task performance, whether the answer is correct; (2) evidence alignment, whether the model is attending to the audio regions required to answer the question; and (3) robustness, whether answers and attributed evidence respond coherently to controlled changes such as event order, overlap, signal-to-noise ratio, or background content. TRACE targets temporal ordering, counting, and event existence, three tasks that remain challenging for current models. We synthesize contextually informed acoustic scenes that allow us to test whether models hallucinate events associated with the context but absent from the audio. This synthesis also allows us to obtain a scalable benchmark for which we know the answers by design. For each scene, we create controlled variants in a programmatic way to stress-test model robustness. We introduce attribution-based metrics that quantify model attention at scale on target events and non-target events. Evaluation across five state-of-the-art models shows that current LALMs often detect where relevant sound evidence occurs, but they do not reliably interpret it, integrate it into the answer, or respond to controlled changes in that evidence. Our benchmark also shows that models still struggle the most with ordering and counting questions.
https://arxiv.org/abs/2610.26894
Drum beats are often conceptualized and produced as arrangements of repeated single-drum sound events or one-shots. Drum Analysis-by-Synthesis (AbS) models combine this structural assumption with reconstruction-based training to perform MIR tasks such as drum transcription, drum source separation, and one-shot extraction, offering a flexible alternative to more traditional task-specific supervision. However, leading AbS approaches struggle to generalize to a broad range of drum timbres (e.g., electronic drum kits), rely on high-quality paired audio-transcription training data, and do not model intra-class timbral variation within a single arrangement (e.g., multiple kick sounds). To overcome these limitations, we propose D-Composer, an autoregressive language modeling framework that decomposes drum recordings into labeled sound events and their onset times. In drum transcription and one-shot extraction tasks, D-Composer performs comparably to or better than a state-of-the-art AbS system and task-specific baselines while generalizing to a broader range of drum timbres. Our results demonstrate that when equipped with abundant synthetic data and compact acoustic token representations, language modeling provides a simple recipe for unifying transcription and sound event decomposition.
https://arxiv.org/abs/2610.22535