Papers

Recognition of domain-specific and low-frequency terms remains challenging for automatic speech recognition (ASR). Although contextual biasing can improve their recognition, directly providing a large terminology dictionary introduces many irrelevant biasing terms. Retrieval-based contextual biasing addresses this issue by selecting candidate terms from an external dictionary, but querying many recognized words requires numerous dictionary lookups and may yield poorly targeted candidates. We propose a training-free contextual ASR framework in which a pretrained speech large language model (SpeechLLM) jointly generates an ASR hypothesis and localizes error spans likely to involve domain-specific terms. Only the localized spans are used to retrieve phonologically similar terms from an external terminology dictionary. The same SpeechLLM then re-recognizes the audio conditioned on the first-pass hypothesis and the retrieved terms, without task-specific model training. To assess applicability across domains, we evaluate the framework on medical, air traffic control, and financial speech. The proposed method substantially reduces dictionary queries while improving the recall and ranking of relevant terminology candidates and second-pass ASR performance across all three domains.

Native full-duplex speech models can respond before a user finishes speaking, making the behavior of the model depend on both the final utterance and when intent-defining information arrives. Existing benchmarks primarily evaluate turn-taking mechanics, whereas safety evaluations typically assume that the complete request is observed prior to response generation. We introduce LateIntent-Bench, a matched-pair benchmark for delayed intent revelation. A shared ambiguous prefix precedes either a benign or a harmful continuation, using a controlled pause to delay when the branches become distinguishable. We define the Premature Response Rate (PRR) to measure whether response onset precedes the revelation of intent. Joint evaluation of harmful and benign engagement distinguishes changes in safety selectivity from general losses in responsiveness. Across 3,136 sessions, four native full-duplex models exhibit distinct response-timing patterns under delayed intent. Three models show increased harmful engagement while maintaining benign responsiveness, whereas one loses engagement with both branches. Inserting a 1.5s silence after intent revelation keeps PRR near the no-pause baseline and substantially reduces changes in harmful engagement. These results demonstrate that evaluating fully specified requests alone fails to capture emerging timing patterns, highlighting the necessity of assessing delayed intent revelation in full-duplex models. The code will be publicly released soon.

Optimal Transport (OT) provides a principled framework for comparing and transforming probability distributions while preserving geometric structure. Recently, OT has gained significant attention in machine learning due to its ability to measure discrepancies between distributions, even when their supports do not overlap, making it effective for tasks such as generative modeling, domain adaptation, and transfer learning. Despite its success in fields such as computer vision and natural language processing, OT remains relatively underexplored in speech research. Speech signals present unique challenges, including temporal dynamics, speaker variability, noise, reverberation, and heterogeneous multimodal representations involving audio, text, and visual information. These factors often lead to distribution mismatches, where OT offers a natural framework for alignment and interpretation. This work aims to promote broader adoption of OT in speech processing by: (1) reviewing OT foundations through intuitive physical interpretations and highlighting connections to modern generative models; (2) presenting computational algorithms suitable for deep learning frameworks; and (3) demonstrating OT applications in cross-domain and cross-modal speech tasks, including speech enhancement, automatic speech recognition, language and speaker recognition, and audio spoof detection. We highlight OT's strong potential for addressing distributional variations in real-world speech applications.

Audio-only pedestrian detection is attractive for urban sensing but limited by weak acoustic cues. An appealing strategy is cross-modal knowledge distillation, in which a video teacher supervises the audio student during training so that the deployed model runs on audio alone. Under this task's severe class imbalance and wide video-audio modality gap, however, what such distillation contributes is unclear. We introduce Trust-Filtered Distillation (TFD), which selectively suppresses teacher supervision on pedestrian samples, and interpret its logit formulation as conditional label smoothing under a shared temperature. Across ten distillation configurations under five-fold cross-session validation on ASPED, most methods yield modest gains in macro accuracy accompanied by small changes in PR-AUC. The main effect is higher no-pedestrian accuracy at the cost of lower pedestrian recall. Adding TFD to logit distillation strengthens this trade-off without improving mean PR-AUC. These findings clarify the benefits and limitations of selective cross-modal supervision by distinguishing operating-point shifts from discrimination gains in imbalanced acoustic detection.

Changes in listener acoustics require active noise control (ANC) filters to be redesigned for new acoustic paths. We introduce PRIME-ANC, a shared neural synthesizer that learns a bounded, path-dependent logmagnitude correction to a regularized path-ratio base. Minimum-phase reconstruction and truncation produce finite-impulse-response (FIR) filters; training optimizes their noise-control performance. Across ten random training/test splits per dataset, PRIME-ANC achieves average held-out one-third-octave reductions of 18.81 and 17.76 dB over 50 Hz-5 kHz on a ten-path dataset and a public earphone database, respectively. On original earphone measurements, it improves upon the path-ratio base by 7.89 dB. Ablation studies support the contributions of both the analytic base and the path-dependent correction. Given calibrated paths for a held-out listener condition, PRIME-ANC generates a path-specific FIR without iterative optimization. With three Gauss-Newton updates, it reaches 21.43 dB reduction, approaching direct weighted least-squares design while producing lower amplification and root-mean-square control output.

Complementarity analyses of multimodal systems commonly report an oracle ceiling (the fraction of examples on which at least one unimodal branch is correct) and interpret the gap between it and realized fusion accuracy as recoverable headroom. We show that this ceiling is not a fusion target. Using frozen self-supervised audio and visual encoders with a trained head, we measure the ceiling and realized gain on CREMA-D under two audio encoders (one whose fine-tuning lineage includes CREMA-D, one clean) and on EAV under the contaminated encoder. Replacing the fine-tuned encoder with a self-supervised one triples the headroom, from +0.055to +0.186. A second clean encoder and a controlled degradation of the contaminated one fall on the same curve, so provenance moves the headroom by changing audio-branch accuracy. Concatenation realizes at most about half of the headroom (converted fraction 0.55), and on EAV the converted fraction is indistinguishable from zero despite larger headroom. A learned router is beaten by plain concatenation everywhere. Oracle headroom measures branch disagreement, not a fusion budget.

Online audio-visual target-speaker extraction aims to remove competing voices while preserving speech quality and bounding lookahead. Existing extractors are built and evaluated for separation on synthetic mixtures, leaving listening quality and meeting behavior largely untested. We introduce Audio-Visual Personalized Voice Quality Enhancement (AV-PVQE), which approaches these requirements from the other direction. We start from a personalized speech enhancement model that reconstructs a requested voice at high quality but confuses the target in 46% of two-speaker mixtures despite clean enrollment. Adding mouth features at its speaker-conditioning input and jointly fine-tuning the visual and reconstruction networks reduces this rate to 1.6%, with no future frames and 20 ms of algorithmic delay. Compared with an online autoregressive audio-visual extractor, AV-PVQE yields separation gains on two synthetic benchmarks and larger gains on recorded meetings, and keeps its advantage on excerpts with more speakers than the fine-tuning mixtures. In personalized P.835 listening tests on two meeting corpora, it improves overall quality over this extractor by 0.57 and 0.63 MOS, with similar mean rating relative to the starting model. Preservation and rejection tests show that it keeps the target intact when no competing voice is present and suppresses competing speech when the target is absent.

Instrument timbre transfer re-voices a performance using the timbre of another instrument. Extracting the target timbre from an audio reference capture more nuances than inferring it from a text prompt. Systems that read timbre from such a clip train a dedicated model for the task, which captures the timbre cleanly but stays a narrow, single-purpose system. More versatile approaches add control to a pretrained music generator, yet a reference clip entangles timbre with genre and melody, so these systems fall back on text to name the timbre. We present MuseTimbre, the first system, to our best knowledge, that transfers timbre from an audio reference to a polyphonic source through conditioning a pretrained music generator. This system employs a multi-pitch estimator to extract pitch information from the source and finetune a CLAP encoder to extract timbre information from the reference audio. Experiments show that across four datasets of real polyphonic recordings, MuseTimbre achieves pitch alignment on par with the baselines while matching the reference timbre far more closely. Results also show that the finetuned CLAP-based timbre extractor is robust to pitch variations, making it useful in timbre similarity measures.

Text-to-music systems are assessed on audio quality and on how faithfully the music follows its prompt, and the CLAP score, the cosine similarity between a music-text model's audio and text embeddings, is the standard objective metric of faithfulness. We ask how accurately that score reflects the text: when an attribute is linked to an instrument (e.g., distorted guitar), does the text embedding capture that binding? To find out, we introduce an attribute swap perturbation: the caption of a real recording is edited by exchanging exactly one property, timbre, lead versus accompaniment, or order of first appearance, between two instruments. We then test four contrastive music-text models and one large audio-language model on whether the audio scores higher against the original caption than against the perturbed one. No contrastive model distinguishes the two captions reliably. The audio-language model does better, but further experiments show that its advantage rests largely on audio-agnostic language priors. Our results thus provide compelling evidence that the CLAP score and related metrics do not capture fine-grained musical meaning or attribute bindings; their representation is closer to a bag-of-words that leaves them insensitive to meaning-changing perturbations of the caption.

Recent progress in speech-driven 3D facial animation has improved vertex-level reconstruction quality, but speech-consistent visible articulation remains difficult. This is because speech production follows structured and constrained articulators' coordination and the mapping from acoustics to motion is inherently one-to-many. Motivated by the structured patterns of visible articulation, we propose a novel articulation-aware framework that models visible speech through directional articulatory motions and composes them into surface-consistent 3D facial motion. To represent visible articulation with three directional articulatory motions, spreading, opening, and protrusion, we propose a Speech–Articulatory Memory (SAM) that captures the correspondence between speech and these motions under phonetic context through retrieval and decoding based on a key-value memory structure. Then, a Topology-aware Articulatory Composition (TAC) integrates the predicted directional articulatory motions under mesh topology to produce surface-consistent 3D facial motion. Experiments on VOCASET and TFHP show that our method achieves state-of-the-art performance on standard reconstruction metrics and improves visible articulatory distance and velocity errors for lip articulation, while a user study confirms clear preference in lip sync and realism.

Audio language models state numbers for acoustic quantities, and neither human opinion nor a judge model says whether such a number is true of the signal. AcoustiClaim extracts each numeric claim from free text, scores it against the instrument that defines the quantity, and classes each quantity by where its reference can be read. Four open-weight systems and one closed model, asked for ten quantities five ways on two corpora, fill 207 cells. Of these, 49 emit fewer than five distinct values, and eight of the 158 cells that can be ranked exceed a rank correlation of 0.3, the bar we set, three with an interval clear of it, five of them one closed model reading pitch. Error sits at or above a constant-predictor floor in every ranked cell but three. The reference decoder we train declines the five voice quantities in prose on 95% of mixtures, with nothing withheld, and states them on the clean twins, reproducing its targets' rule from audio alone. With a calibrated threshold, withholding lowers error on all ten quantities on the mixtures in the mean and on eight at every split, against at most 0.6% from a random selector. A linear baseline orders errors at least as well as ours. F0 s.d. and shimmer stay above the constant floor.

Target-speaker automatic speech recognition (TS-ASR) must identify and transcribe a desired speaker under varying overlap and noise conditions. These changes alter the acoustic evidence for the target speaker in the speech mixture, motivating inference-time calibration of speaker conditioning. We introduce asymmetric classifier-free guidance (CFG) for TS-ASR using Whisper: the speaker-conditioned branch predicts the target transcript, while the speaker-unconditioned branch predicts serialized multi-speaker transcripts. CFG adjusts the contribution of speaker conditioning during decoding through a single guidance scale. We select a global guidance scale on target-domain development data and train a lightweight encoder-based predictor to adjust it for each utterance, keeping the recognition model fixed. Under domain shifts, our full system achieves relative word error rate (WER) reductions of up to 21.8% over the condition-only baseline, and 5.6% over standard conditional decoding of the same CFG-trained model. Oracle analysis shows that substantially larger WER reductions are possible through utterance-level scale selection and identifies how beneficial adjustments vary with domain shifts.

We introduce target-speaker unlearning ASR (TSU-ASR) task in a fully end-to-end framework for multi-speaker ASR and diarization. Given a multi-speaker utterance and a set of opt-out speakers who do not wish to have their speech transcribed, the task requires an ASR system to transcribe all speakers except the opt-out ones, while still indicating when those speakers are active. As a first step towards tackling this task, we introduce a novel, light-weight Enrollment-Conditioned Gating (ECG) module attachable to a frozen dual-stream speech LLM that enables ASR for new opt-out speakers dynamically during inference, even those who were not seen during initial ECG training phase. Our experiments on both AMI (English) and AliMeeting (Mandarin) datasets show that speech transcription accuracy for corresponding opt-out words or characters falls from 72.3% to 48.2% and from 73.6% to 27.3%, respectively, while retained speakers' transcription error rates maintain more or less the same. Our approach provides a practical solution for modern video conferencing platforms, allowing speakers to dynamically opt-out from automated AI transcriptions without forcefully leaving the meeting sessions, enabling a privacy-preserving interface for potentially millions of online meetings daily.

Signal-level speech privacy mechanisms suppress linguistic content while preserving acoustic information needed by downstream sensing applications. However, their privacy settings are typically evaluated/selected offline and remain fixed during deployment, even though speech-content leakage can vary substantially across utterances and speakers. Adapting privacy protection at runtime requires estimating how much speech remains recoverable, but conventional measures such as WER or PER require ground-truth transcripts and therefore cannot be computed online. We present CLEAR, a reference-free approach for estimating speech-content leakage at runtime using disagreement among heterogeneous ASR systems. Our key insight is that independently trained ASRs exhibit consistent behavior when linguistic content remains recoverable and increasingly disagree as privacy transformations obscure speech. Using configurable speech-suppression mechanism, we show that cross-ASR disagreement closely tracks transcript-grounded leakage across privacy operating points, achieving a correlation of 0.8. We further characterize the latency-accuracy trade-off of heterogeneous ASR subsets and use their hypotheses to identify potentially exposed words. These capabilities enable privacy to be treated as a runtime property rather than a fixed configuration: CLEAR can communicate residual speech exposure to users and provide feedback for dynamically adjusting privacy aggressiveness while retaining acoustic utility for downstream sensing tasks.

Automatic speech recognition is typically trained assuming that the reference transcript is the only valid labeling of an utterance, yet even nominally verbatim transcripts contain localized differences in pronunciation, spelling, or lexical realization that the acoustics do not uniquely determine. Omni-temporal Classification (OTC) tolerates such noise by adding wildcard paths to the connectionist temporal classification (CTC) alignment graph, but its word-level arcs are too coarse, since bypassing one unsupported token discards supervision for the whole word. We move wildcard arcs to token granularity so unsupported tokens can be bypassed while the rest of the word stays supervised, and we combine token- and word-level arcs as complementary escape paths. Across 19 languages and three corpora, token-level OTC improves over CTC on all 25 tasks. We also replace epoch-indexed relaxation of the wildcard weights with a predictive-entropy-indexed schedule, which performs comparably while reducing dependence on training length. Combining this schedule with the hybrid graph gives the lowest mean word error rate (WER) on every corpus and a 9.45% average relative WER reduction over CTC. Independent validator transcriptions show that token-level models place significantly more wildcard-bypass probability than CTC on disputed characters, indicating that token-level tolerance targets localized transcript ambiguity.

Open-vocabulary Sound Event Detection detects and temporally localizes acoustic events described by arbitrary text queries. Progress in this emerging field is hard to assess: recent methods report on disjoint task subsets under incompatible protocols without a benchmark spanning the acoustic domains and query types the task presents. We establish a comprehensive benchmark by assembling six temporally-annotated tasks: four with fixed class vocabularies over domestic, urban and mixed indoor/outdoor scenes, plus two free-text grounding tasks. We evaluate five recent methods on identical data and metrics under a label-space zero-shot criterion. Our benchmark demonstrates that no prior method is competitive across all six tasks. We then introduce COSED, which surpasses prior work on five out of six tasks while staying on par with the best method on the sixth, with margins of 12-33% on three of them. COSED is the only system in our comparison competitive on every task, and so generalizes across acoustic domains and query types better than prior work. We also provide a leave-one-out ablation study that isolates the sources of the performance benefits: scoping negatives to their corpus of origin (25.8%), combining closed- and open-world supervision (16.8%), and improving temporal processing (16.4%).

Phoneticians often need to construct stimuli in which specific acoustic cues are precisely manipulated while preserving decent speech quality. Classical synthesis and modern neural methods sit along a trade-off between precise parametric control and high fidelity, and neural synthesis typically demands more data than phoneticians can easily obtain. We present ARIS (Analytic Resonant Interpretable Synthesis), a neural source-filter model that pairs neural parameter estimation with deterministic DSP synthesis. Every control is a coefficient of the synthesizer, so F0, formants and the glottal source can be edited directly. On five small single-speaker corpora in three languages, ARIS resynthesizes speech with quality comparable to WORLD and edits single parameters more accurately than Praat KlattGrid, with negligible crosstalk between cues. Compared with HiFi-Glot, pre-trained on a large corpus and fine-tuned on the same data, ARIS scores slightly lower on predicted naturalness but reproduces the recordings more faithfully and manipulates them more precisely. Audio samples: https://n1r.github.io/ARIS_nsf/.

Deep learning-based speech enhancement is increasingly deployed on-device in hearing aids, headsets, and earbuds. Most of these devices, however, can only accelerate static int8 graphs, so a depth-varying network must be implemented as several graphs, orchestrated by a policy. In this paper, we supervise every intermediate depth of one causal model, then we fine-tune its output heads to guarantee that deeper outputs are never worse than shallower ones. Using this training protocol, we can derive a family of static models that are more Pareto-efficient than their equivalently-sized counterparts trained from scratch on the same budget. Specifically, we achieve up to 0.11 higher PESQ for equivalent compute, and match the best PESQ at 30% less compute. We then quantize the models to int8 and measure the latency-quality frontier on an STM32N6 microcontroller. On VoiceBank-DEMAND, the dynamic enhancer lies on the same frontier as the static models, rather than trading quality for dynamic execution. Running the policy on the companion Cortex-M55 takes only 26 $\mu$s per frame, while splitting the enhancer into separate NPU graphs adds 2.2% latency overhead. The cost of dynamic execution is therefore small.

Deploying real-time speech enhancement on resource-constrained devices requires meeting strict latency, memory, and energy constraints. Microcontroller NPUs can accelerate neural inference under these constraints, but only through a restricted set of operators in static, integer-quantized graphs. Recent speech-enhancement networks have reduced parameter counts and MACs to levels nominally suitable for microcontrollers, but their operators and execution patterns often remain incompatible with restricted NPUs. We address this gap by redesigning LiSenNet, a 37k parameter sub-band dual-path model, for the STM32N6570-DK Neural-ART accelerator. We replace its recurrent bottleneck with convolutional frequency and temporal mixers, reformulate unsupported operations as static int8-compatible primitives, and use bounded decoder activations to preserve quality after quantization. On VoiceBank-DEMAND, the final NPU-compatible model matches or exceeds the recurrent LiSenNet baseline, reaching PESQ 3.08 versus 3.01 in FP32 and 3.01 versus 2.93 in int8. Deployed on a microcontroller, it processes each 16 ms input hop in 4.83 ms, corresponding to a real-time factor of 0.30. Stateless receptive-field recomputation is an order of magnitude slower at the same frame rate despite higher accelerator utilization. These results show that parameter count and operator compatibility, quantization range, and persistent streaming state must be co-designed to achieve efficient real-time speech enhancement on restricted NPUs.

Audio classification is inherently a multi-label task, as real-world acoustic environments contain multiple simultaneous sound events. When new sound classes emerge, models must incorporate them without forgetting previously learned ones: a challenge known as class-incremental learning. Existing methods rely on storing past data and iterative gradient updates, which struggle under incomplete multi-label supervision because only the newly introduced classes are annotated at each phase, leaving old-class labels unavailable. We investigate exemplar-free analytic continual learning as a principled alternative, in which a linear classifier is updated in closed form without storing historical recordings or performing incremental back-propagation, and previously learned weights remain intact by construction. Building on analytic learning, we further propose ALMA, which addresses incomplete supervision and class imbalance through continuous old-class score estimates and frequency-based sample weighting. Experiments on a 50-class AudioSet-R benchmark across three incremental setups show that the analytic learner substantially outperforms gradient-based methods, and previously learned classes retain nearly unchanged detection performance as new classes are added. This study shows that ALMA is a simple yet effective solution to multi-label audio class-incremental learning.

We study how to organize Transformer depth through recurrence in flow-matching text-to-speech, varying the amount, order, and place?ment of weight reuse. Seven layouts perform 18 block calls per network evaluation under a common training objective and sam?pler. On Seed-TTS and LibriSpeech-PC, SEQUENCE applies each of nine blocks twice consecutively, retaining competitive intelligibil?ity, speaker similarity, and predicted speech quality at 32 sampling steps with 47.1% fewer parameters than the unshared baseline. Cy?cling six blocks three times further reduces model size but raises 32-step word error rates relative to cycling nine blocks twice. At matched parameter counts and executed depth, reuse order and shar?ing position produce different quality trade-offs. These comparisons depend on sampling budget: Prefix and Suffix have similar 32-step word error rates, but Suffix is worse by 3.44 and 5.97 percentage points at four steps on the two datasets, respectively. Only Middle ranks first or second in mean word error rate at 32 and four steps on both datasets. These results show that the organization of recurrent computation affects synthesis quality, and that reuse layouts should be selected for both the sampling budget and the quality dimensions of interest.

Cross-speaker acoustic-to-articulatory inversion requires accounting for anatomical differences between speakers. We propose a geometric adaptation framework that uses anatomical landmarks, primarily on vertebrae and dental structures,to transfer predictions from a fixed inversion model to unseen speakers. An affine transformation followed by thin-plate spline (TPS) deformation maps the predicted contours of 10 vocal-tract structures into each target speaker's geometry without retraining. Landmarks are identified in one selected /u/ frame per speaker as a common phonetic reference without assuming identical articulatory configurations across speakers, and the resulting mapping is reused across recordings. We train the model on a single-speaker rt-MRI database and evaluate adaptation on eight speakers from a separate multi-speaker rt-MRI database. We compare affine and TPS configurations using 12 or 14 landmarks. Affine12+TPS14 achieves the lowest mean point-to-closest-point error of 3.19mm. These results support the combined value of anatomical landmark information and nonrigid alignment.

Just as motorisation made road traffic a defining noise source of the twentieth-century city, low-altitude aircraft may reshape urban acoustic exposure in the twenty-first. Yet how this noise interacts with existing sound environments and three-dimensional urban form across global cities remains unclear. We modelled identical low-altitude aircraft operations across ten urban districts spanning all inhabited continents. Horizontal exposure varied markedly across and within cities, depending on road noise conditions and urban morphology, which also strongly influenced overall attenuation. Some semi-enclosed spaces with low noise levels before low-altitude aircraft operations experienced increases above 20 dB(A), creating pronounced local contrasts in urban noise exposure. Vertically, exposure varied with height and acoustic visibility to the flight route, creating vertical exposure inequalities between storeys, façades and buildings at comparable heights. These findings show that low-altitude aircraft can reshape noise exposure across global cities, requiring route assessment to distinguish newly exposed from already exposed areas and to account for three-dimensional exposure.

Although room impulse responses (RIRs) encode source-distance cues, conventional monaural source separation focuses on recovering audio content without estimating source-specific RIRs, losing the associated spatial information. To address this limitation, we propose Distance-Aware Monaural Source Separation using Multi-RIR Estimation (DAMSEP), the first end-to-end framework that is jointly trained for source separation and multi-source RIR estimation from a single-microphone mixture. DAMSEP integrates a separation backbone with shared dereverberation and RIR estimation modules to jointly recover clean sources and source-specific complex convolutive transfer functions under source estimation and reverberant reconstruction objectives, enabling relative near/far ordering through the direct-to-reverberant ratios of the corresponding RIRs. For comprehensive evaluation, we introduce HETMIXR, which spans heterogeneous source content and diverse simulated room conditions with source-specific RIRs and geometric distance annotations. Experiments on HETMIXR demonstrate superior performance in source separation, RIR estimation, and distance ordering. Ablation studies reveal the complementary benefits of source supervision and reverberant reconstruction, while additional evaluations show generalization to single-speaker inputs and mixtures generated using measured RIRs from an unseen room. Our code and dataset are available at https://github.com/Wenanzhi/DAMSEP.

Speech enhancement systems are often developed for a fixed sampling rate, while time-frequency models become more expensive as the number of frequency bins increases. We propose TF-Refiner, a sampling-frequency-independent model that decouples the deep analysis bandwidth from the full-band input and output. A deep encoder processes the band below a configurable cutoff, while a shallow decoder combines the encoded features with input-dependent high-band queries and predicts local complex filters applied to the original noisy STFT. A single parameter set trained at 16 and 48 kHz is evaluated at various sampling rates. On VoiceBank+DEMAND, the universal model outperforms the rate-specific counterparts in PESQ, STOI, and log-spectral distance across the evaluated rates, including rates unseen in training. Random-cutoff training enables inference-time selection of cost-quality operating points without retraining or changing the output bandwidth. These results support configurable analysis bandwidth as a practical design choice for multi-rate full-band enhancement.

Voice agents must complete users' tasks despite noise, reverberation, and competing speech. Evaluating agents' robustness therefore requires following how acoustic conditions affect the conversation and the actions taken on the user's behalf. This overview examines what existing benchmarks reveal about agents' ability to complete tasks under acoustic stress and where further task-based evaluation is required. We then introduce TRACE, a practical workflow for designing, running, and interpreting evaluations of acoustic robustness in task-oriented human-agent interactions: the same agent attempts a specified task with an original recording and an acoustically stressed copy, and the resulting conversations are scored for task completion, wrong actions, recovery, and user effort. Finally, we explain how results from these evaluations can guide changes to an agent to prevent wrong actions and improve recovery.

We adapt Reinforce Adjoint Matching (RAM), a reward-based post-training method, to generative speech enhancement (SE). Starting from a pretrained SE model, RAM tilts the model's conditional distribution toward outputs with higher reward. During training, the current model generates enhanced speech on-policy, evaluates each generated endpoint with a potentially non-differentiable reward, and analytically re-noises the endpoint to construct inputs for a reward-guided regression objective. This enables post-training directly on real recordings using weak supervision, such as text transcripts, without requiring paired clean speech targets or reward gradients. We investigate word error rate (WER)-based post-training and whether recognition performance can be improved without compromising perceptual speech quality. Experiments on real CHiME-4 recordings reduce WER by 5.08 percentage points relative to pretrained FlowSE without reducing any of the reported non-intrusive speech quality metrics. A subjective listening test at the default reward scale finds no statistically significant preference between the post-trained and pretrained models.

The acoustic front-end determines which forensic cues a speech deepfake detector can exploit. The wavelet scattering transform (WST) provides stable multiscale coefficients with explicit coordinates, yet direct flattening obscures the parent relation between paths. We introduce WST-Graph, reconstructing these paths as a sparse modulation-carrier grid for an AASIST graph backend. Modulation-level normalization and length-aware adaptive local attention pooling produce fixed relative-time representations while retaining the acoustic axes before learned adaptation. This yields a waveform-to-graph interface with a fixed, parameter-free WST. Our configurations remain competitive with AASIST while using approximately 60% fewer trainable parameters and show clear gains on selected out-of-domain benchmarks. These results underscore the value of preserving parent-child relations within the carrier-modulation topology when constructing a compact, physically grounded interface for graph-based speech deepfake detection. Code will be released at https://github.com/saki-ciallo/wst-graph.

Target speech extraction (TSE) typically trains a separate extractor per cue, and visual-cue systems often need corruption-matched training to remain robust under visual frame corruption. We show that one autoregressive LLM backbone, TSE-Omni, can serve both temporally synchronous cues (lip movements, co-speech gestures) and asynchronous cues (enrollment audio, text). TSE-Omni is driven by next-token prediction: each step predicts target speech semantic tokens from its own past outputs, which we term self-enrollment, forming a continuous target-speech context initialized by the enrollment cue (asynchronous audio or text, or a short visual prefix). This enables audio-visual compensation: the model uses synchronized visuals when intact and its token history when visual frames are missing. Under clean visuals, TSE-Omni matches strong discriminative and generative baselines (SpeechBERTScore 0.81 on VoxCeleb2 and 0.89 on LRS3 zero-shot) with higher DNSMOS. On the same VoxCeleb2 test set, after a 2 s clean visual start, removing the remaining visual frames leaves SpeechBERTScore at 0.81. It remains usable under sparse overlap and multi-speaker interference, and supports streaming inference. Project page: https://alexwxwu.github.io/tseomni-main/.

Speaker distance estimators are trained almost exclusively on simulated room acoustics, because real recordings annotated with the true talker-to-microphone distance are scarce. We show that models trained this way transfer poorly. On three real corpora we evaluate, simply predicting the average distance of the corpus is more accurate than any learned model. Then, we ask how few labelled real utterances are needed to make a frozen, synthetic-trained estimator useful, and study post-hoc calibration maps that rescale its output without gradients or retraining. An analysis of the achievable error shows that what the calibration is not limited by the absolute accuracy of the estimator, but how well it orders utterances by distance, since a constant bias or a wrong output scale is removed exactly by the calibration itself. Balancing this against the cost of estimating each coefficient from few samples yields a criterion that accounts for which map wins on which corpus and at which annotation budget, together with a shrinkage variant that requires no hard decision. Our findings suggest selecting synthetic checkpoints by linear correlation with true distances rather than by absolute error. Code, datasets, and analysis are available at https://github.com/michaelneri/audio-distance-estimation.