Papers

Autoregressive Transformer and Conformer models have achieved state-of-the-art performance in automatic speech recognition (ASR). However, their sequential token generation mechanism imposes a significant bottleneck during decoding, resulting in substantial inference latency. While speculative decoding has emerged as a promising lossless acceleration technique, existing ASR-oriented approaches either suffer from low draft acceptance rates due to suboptimal draft quality or rely on specially designed external draft models, which limit their generality and practicality. To overcome these limitations, we propose EAGLE-ASR, a novel framework that accelerates ASR inference via EAGLE-style speculative decoding while strictly preserving lossless output quality. Specifically, we introduce Speech-Draft, a lightweight autoregressive module that serves as a universal draft model by jointly leveraging encoder acoustic features and hierarchical decoder representations from the target ASR model. Furthermore, we adopt a multi-step generation strategy with self-distilled data to train the Speech-Draft, enabling it to better approximate the target model's output distribution and maintain causal consistency during multi-token generation. Extensive experiments on representative Whisper and Conformer AED models demonstrate that EAGLE-ASR achieves significant lossless speedups of up to $1.9\times$ and $2.4\times$, respectively, consistently outperforming existing speculative decoding baselines for ASR. Our proposed framework further generalizes effectively to multilingual ASR and speech translation tasks, establishing EAGLE-ASR as a practical and universally applicable solution for accelerating modern ASR systems without compromising recognition accuracy.

Target speaker extraction (TSE) is a key frontend for enabling audio large language models (LLMs) to recognize and understand a designated speaker in overlapping speech. However, multilingual mixtures exhibit heterogeneous extraction demands, causing fixed-depth systems to either under-process challenging inputs or waste computation on inputs that have already converged. Existing TSE frontends generally apply uniform computation, while Audio-LLM adaptation methods neither explicitly model language-conditioned extraction demands nor preserve plug-and-play compatibility with frozen downstream models. We therefore propose LATTE, a lightweight 5.2M-parameter frontend that combines an acoustic Difficulty Estimator, a task-adaptive phonology-informed prior learned from multi-depth TSE gains, and an Adaptive Hourglass Extractor to route each mixture without requiring language identity at inference. On PolyExtract, LATTE achieves a macro SI-SNRi of 15.07 dB at 0.30 GFLOPs, retaining performance within 0.09 dB of a 0.53-GFLOP fixed-depth model while reducing computation by 43.4%, and lowers the average WER from 16.8% to 14.6% relative to TF-GridNet. By jointly modeling sample-level extraction demands and phonological structure, LATTE provides a practical framework for quality-aware, compute-efficient multilingual TSE and offers a scalable path toward robust Audio-LLM deployment.

Speech recognition on low-cost microcontrollers is, in practice, limited to keyword spotting or a fixed list of commands. We present Oído, an inference engine that runs open-vocabulary English speech recognition entirely on an ESP32-S3 (240 MHz dual-core Xtensa LX7, 512 KB of SRAM, 8 MB of PSRAM, 16 MB of flash, no neural accelerator). Oído executes a 13M-parameter Conformer-CTC model with int8 weights, per-frame dynamic int8 activations and int8 attention on the chip's SIMD instructions, with a schedule designed around flash and PSRAM bandwidth. On LibriSpeech test-clean/test-other the on-device arithmetic reaches 3.70%/8.23% word error rate (WER), within 0.12 points of full precision, and 3.33%/7.15% with an on-chip 1.3M-parameter language model; a transducer head gives 2.99%/6.67% and an int4 variant fits in 8.3 MB. This is 2.3-2.8x fewer errors than the vendor's recognizer on the same chip, and fewer than Whisper tiny.en and Moonshine tiny in full precision on a laptop, including in real noise and reverberation. Measured on a physical board, the firmware runs at 1.97x the duration of the audio on one core, with transcripts identical to an instruction-exact emulator: accurate, but not yet real time (a 4 s command is transcribed about 8.7 s after it ends). The two-core schedule gives wrong results on silicon and is disabled. A streaming variant shows partial text while the speaker talks and delivers the final text 4.1-11.1 s after clips of 3.6-10.6 s (7.2-22.4 s in utterance mode) at 4.86%/11.03% WER; a Spanish model makes 34-60% fewer errors than multilingual Whisper tiny. Code and models are released.

Recent accent classification systems can achieve high accuracy in standard in-corpus evaluations. However, when speakers overlap between the training and test sets, such results may overestimate model robustness compared with real-world deployment scenarios. This study revisits the robustness of DistilHuBERT-based accent classification by using L2-ARCTIC and multiple external English accent datasets. The comparison covers file-level and speaker-disjoint splits, along with balanced external-data training and teacher-student pseudo-labeling. The large performance gap between in-corpus and external evaluations suggests that file-level evaluation substantially overestimates performance, while speaker-disjoint evaluation can provide a more realistic estimate of the generalization to unseen speakers. In this study, models trained on external data alone do not reliably generalize to L2-ARCTIC. Among the tested configurations, balanced external-data training offers the most promising supervised setting. The pseudo-labeling experiments further indicate that confidence-based filtering alone is insufficient under domain mismatch. Overall, this study suggests that robust accent classification evaluation should simultaneously include speaker-disjoint evaluation and cross-corpus testing.

Speech anonymization is evaluated with automatic speaker verification attackers whose performance depends on the attribute distribution of both their training and evaluation data, which may bias privacy estimates for some speaker groups. To alleviate this issue, we introduce VoxAnon, a dataset built from VoxCeleb1 and 2 with training and evaluation splits balanced across gender, age, accent, and pitch. We extract and cross-check attributes using two complementary methods, balance trial lists per subgroup, and draw test/enrollment utterances from distinct video sources in order to limit linguistic and session bias. Comparing VoxAnon with unbalanced training and test data across four attackers and three anonymizers, we show that attribute imbalance changes which subgroups appear hardest to re-identify, in directions that depend on the attribute. In the balanced setting, residual disparities persist, mostly tied to accent and pitch, and depend on the anonymizer/attacker pair. Code and trial lists are publicly available.

Modeling human reactions to spatial audio requires understanding both what is heard and where it comes from. Prior work generates motion offline from complete acoustic and spatial input sequences. Humans, however, react to sounds without knowing what they will hear next. We present Sonic-AMG, to our knowledge the first framework for streaming human motion generation from spatial audio. The model encodes acoustic content and body-relative source geometry as separate tokens. Conditioned on these tokens, a causal transformer with a flow head generates continuous motion latents, from which a causal decoder reconstructs motion segments. Standard teacher-forced training derives motion histories and body-relative source locations from ground-truth motion, whereas streaming inference relies on generated motion for both, resulting in a training-inference mismatch. To reduce this history discrepancy, we introduce rehearsal training, which combines full-sequence teacher-forced supervision with supervised prediction along short autoregressive self-rollouts. To keep spatial conditioning consistent with generated motion, we propose spatial feedback, which recomputes body-relative source locations from generated root poses during both rehearsal training and inference. On SAM, Sonic-AMG achieves state-of-the-art FID and R-precision. It further supports real-time online generation of coherent reactions across sound events and source locations. Project page: https://jingchengni.com/sonic-amg/.

This paper describes our systems for the Track 1 time-stamped speaker-attributed ASR task of the SmartGlasses Challenge. To handle overlapping speech, background noise and far-field conditions in two-person conversations, we employ SoulX-Transcriber, a multimodal large language model that takes full-length audio as input and generates structured transcriptions. In addition, we investigate the effect of speech frontend processing, including a rule-based Guided Source Separation method and a data-driven speech separation model, both of which improve performance on the validation set. Considering inference stability, our final submission uses the fine-tuned SoulX-Transcriber system, achieving a tcpCER of 6.22 on the test set.

Open audio-language models are used as judges of generated music: given a clip and its text prompt, the model rates the clip, and the judge is validated by how well its ratings correlate with expert ratings. The correlation does not show what the judge responds to: the music, the text prompt, or the way the question is asked. We propose four probes that test each of these, three of them read against a control computed from the human ratings: removing the audio, replacing the clip with another model's clip for the same prompt, swapping the order of two clips in pairwise judging, and rewording the instruction. We apply them to five open judges on all 2,748 clips of MusicEval, and the last two to the SongEval songs. Without audio, no judge exceeds what the prompt alone predicts from the human ratings, and for prompt alignment every judge falls below it. With audio, every judge follows the clip it hears when rating prompt alignment, two do when rating musical quality, and the strongest judge's quality score tracks the experts' alignment rating more than their quality rating. The pairwise verdicts of the four smaller judges follow the position of a clip in the pair (its slot) rather than its quality on both rating sets. For musical quality, rewording the instruction moves the correlation of most judges as much as changing the judge. MusicEval's prompt assignment bounds any prompt-only judge at the model level. We recommend reporting the four probes with every music judge.

We study query-conditioned multi-instance open-vocabulary sound event localization (QC-MI-OVSEL), where a system must recover all active physical sources matching a natural-language category query, including when the queried category is absent. We introduce the QC-MI-OVSEL benchmark, a 40,000-scene first-order Ambisonics benchmark with 80,000 balanced present/absent query views over 83 classes, including 17 novel classes excluded from training and development. Each frame is annotated with the activity and 3-D direction of every query-matching source, with up to three simultaneous targets. We further propose Q-RSet, which uses query-conditioned source slots and responsibility-based evidence aggregation for source-set prediction. Experiments show improved source recovery, localization recall, cardinality estimation, exact-set recovery, and absent-query suppression over complementary baselines. Analysis indicates that novel categories, paraphrased queries, and higher source multiplicity mainly impair source discovery and set completion, while directional accuracy remains comparatively stable for successfully recovered sources.

Recent advances in speech synthesis technology have enabled the generation of speech with diverse characteristics. One such characteristic is a foreign accent. In our preliminary experiments, we found that UTMOS, a widely used automatic speech quality assessment system, tended to assign lower scores to speech with stronger foreign accents, even though it is produced by a human speaker. This observation raised a question: How do humans evaluate the naturalness of foreign-accented speech? We conducted a large-scale subjective evaluation involving both native and non-native listeners under different evaluation instructions. The results showed that humans also considered accentedness when evaluating naturalness, but that defining naturalness in terms of human-likeness and selecting raters who have greater exposure to the target accent may enable more appropriate evaluation. This study aims to provide a starting point for future discussions on fairer and more appropriate evaluation of synthetic speech with a foreign accent.

Large audio-language models (LALMs) frequently exhibit object hallucination, affirming absent sounds. The dominant mitigation, Audio-Aware Decoding (AAD), contrasts logits under real audio against an all-zero "blank" waveform, flagged by prior review as out-of-distribution. We introduce a black-box causal test that isolates the claimed evidence itself: localize a claim's evidence region via CLAP similarity, remove it by spectral gating, and compare against a random-window control across a 2x2x2 grid on three LALMs. On Qwen2-Audio-7B, localized ablation's raw AUROC trails blank's (0.613 vs. 0.721); this is a duration confound, not a localization failure, since blank removes the entire clip versus the grid's 2-4 s. Holding duration fixed, localization beats random removal at every cell on both models. Raw detection value remains model-dependent, especially on Audio Flamingo 3. The same evidence corrects decoding: contrastive decoding combining blank and language-prior contrasts, gated and ensembled, raises Qwen2-Audio accuracy from 57.1% to 71.9%, beating AAD alone. A unified causal-grounding profile does not transfer between models, indicating different grounding regimes. The source code and a detailed report of our findings, including identified limitations, are available at https://github.com/khalooei/causa.

Broadcast archives offer scale and diversity for self-supervised learning (SSL) but contain substantial noisy, non-speech, and multilingual content. Prior work typically relies on voice activity detection or language identification (LID) to extract speech, ensuring presence rather than quality. Whether and how to filter unlabeled speech by quality for SSL, particularly from noisy broadcasts, remains underexplored. We present a large-scale Finnish SSL study starting from 2.1M hours of unconstrained broadcast audio. We propose automatic speech recognition (ASR) confidence as a proxy for pretraining data quality and show 3-7% relative WER improvements over LID-based filtering despite a smaller training set. We further address the lack of controlled architectural comparisons in speech SSL by evaluating Transformer, Macaron-Net, Conformer, and E-Branchformer at matched parameter budgets. Our models improve upon prior Finnish SSL baselines, particularly at Base scale and out-of-domain.

Joint audio-video diffusion models generate synchronized content, but editing existing clips requires target changes without disrupting non-target structure or cross-modal correspondence. We present AV-FlowEdit, a training-free and inversion-free framework for prompt-driven audio-video editing. Our key insight is that source–target flow differences provide both cross-modal editing directions and preservation cues. Under shared perturbations, they jointly update the video and audio branches, supporting visual-only, audio-only, and joint edits while retaining pretrained cross-modal interactions. To suppress drift accumulated during sampling, Cross-Modal Event-Aware Spectral Trajectory Preservation combines flow saliency, aligned audio activity, multiscale low-frequency displacement, and diffusion-stage scheduling. This adaptively preserves unsupported regions while allowing intended audio-visual events. AV-FlowEdit requires no model training, external masks, or optimization-based inversion. Experiments show strong prompt alignment and temporal consistency, together with competitive audio quality and cross-modal coherence.

Recently, generative neural synthesizers have enabled zero-shot instrument cloning from reference recordings and timbre generation from text prompts. However, existing systems generate complete audio segments offline, precluding their use as playable instruments. We present LiveSynth, a streaming neural synthesizer with flexible timbre control and interactive generation capabilities, including zero-shot instrument cloning and text-to-instrument synthesis. A feedback-free causal Transformer conditioned on MIDI events and a CLAP timbre embedding outputs continuous audio latents at 100 Hz, which a causal variational autoencoder decodes into audio. Frame-wise timbre conditioning enables continuous timbre morphing, while MIDI-conditioning dropout enables autonomous continuation from a performed prefix. On the NSynth and Slakh datasets, LiveSynth achieved timbre similarity and spectral reconstruction quality comparable to offline baselines. On an M3 MacBook Air, LiveSynth required 8.8 ms of computation per 10 ms audio frame (an RTF of 0.88) with 10 ms algorithmic latency, enabling real-time interaction with a performer.

Recent zero-shot Text-to-Speech (TTS) systems can clone unseen speakers, yet controllable expressiveness remains limited by a gap between semantic intent and acoustic realization. Text descriptions offer a natural control interface but provide weak acoustic grounding, whereas speech style prompts provide concrete prosody but entangle style with the prompt speaker's timbre. To bridge this gap, we propose E3-System, a unified intent-to-speech framework for expressive zero-shot TTS that couples three requirements: Expressive style modeling, Explicit timbre-style control, and Easy-to-use natural-language interaction. Given a target speaker and either a natural-language style description or a speech style prompt, E3-System generates expressive speech while preserving the target timbre. The key idea is to close the full control loop: building a large expressive style space, grounding user intent into usable acoustic style references, and synthesizing speech with disentangled timbre and style. To realize this loop, EmoData provides a thousand-hour multilingual emotional speech dataset, Style-RAG grounds text descriptions into high-quality acoustic style prompts from these data, and Dual-TTS transfers the grounded style to arbitrary target speakers through VC-augmented timbre-style disentanglement. Experiments show that E3-System achieves strong speech quality, speaker preservation, and explicit expressive controllability.

Diffusion-based TTS has demonstrated strong capability in emotional speech generation. However, existing emotion alignment methods primarily reinforce target emotional expressions while overlooking modeling of competing emotional modes, limiting fine-grained controllability. This issue becomes more pronounced under classifier-free guidance (CFG), which relies on contrasting conditional trajectories to shape generation. We propose Emotion Bidirectional Preference Optimization (Emo-BPO), a simple effective framework that jointly learns emotion-aligned and emotion-contrastive score functions from reordered same-text emotional pairs. Our method requires neither auxiliary reward models nor additional annotations, and can be seamlessly integrated into existing diffusion-based TTS systems to improve emotional controllability and expressiveness.

Embodied Audio-Visual Navigation (AVN) requires an agent to locate sound targets by jointly leveraging visual and auditory cues. While prior work has achieved notable success with static sound sources, real-world scenarios often involve moving targets, giving rise to the more challenging problem of Dynamic Audio-Visual Navigation (D-AVN). Existing approaches largely rely on semantic auditory features, which limits generalization to unheard sound categories and hinders fine-grained spatial reasoning. To address these challenges, we propose a novel Two-Stage Multi-modal Learning framework for D-AVN. In the Physical Alignment stage, the proposed Intensity-Calibrated Geometry Encoder (ICGE) performs physics-informed alignment of sound intensity with visual geometry to capture category-agnostic spatial correspondence. In the Scene Aggregation stage, a Transformer-based Cross-Modal Scene Aggregator (CMSA) dynamically integrates multi-modal features to construct a robust state representation for policy learning. Extensive experiments on Replica and Matterport3D demonstrate that our method achieves state-of-the-art success rates, while significantly improving generalization and navigation efficiency in complex environments and under unheard sound scenarios.

Reward-guided inference-time alignment can improve joint audio-visual generation, but optimizing multiple alignment objectives requires many expensive model evaluations, making existing methods costly under limited compute. Particle-based search enables global exploration but incurs estimation error with a finite population, while Metropolis-Hastings (MH) can refine candidates locally but benefits from a strong initialization. We introduce Spend Where It Matters (\modelname), a two-stage, training-free, reward-agnostic framework that combines particle search with MH refinement under a fixed compute budget. In the first stage, we adaptively allocate particles across denoising checkpoints using online variance estimates, minimizing estimation variance while concentrating compute where it is most useful. This search provides a promising candidate for MH refinement, which uses local noise-and-denoise proposals on completed outputs to further improve reward. By starting from a strong candidate, MH can recover additional reward without requiring a larger particle population or costly global exploration. We show that, at matched compute, combining particle search with MH refinement achieves higher rewards than either approach alone. On JavisBench-Mini, at closely matched compute, \modelname achieves mean relative gains across seven metrics of $30.7\%$ over BoN and $14.1\%$ over EvoSearch with JavisDiT++, and $9.2\%$ over BoN and $17.0\%$ over EvoSearch with LTX-2. Beyond joint audio-visual generation, we also demonstrate the benefits of adaptive particle allocation on SD1.5 image generation, improving AestheticScore by $2.2\%$ and ImageReward by $6.0\%$ over uniform allocation.

Audio-visual speech enhancement (AVSE) improves speech intelligibility in noisy environments but often relies on computationally expensive models unsuitable for resource-constrained edge devices. This paper introduces a reinforcement learning-driven neural architecture search (RL-NAS) framework to discover compact, efficient AVSE architectures. The search space preserves temporal alignment between audio and visual streams, while a novel reward function jointly optimises speech enhancement performance, measured by SI-SNR improvement, and model compactness. Using proximal policy optimisation with constraint-aware action masking, the framework evaluates only 35 candidate architectures and identifies a model with 2.9 million parameters, compared with 22 million in the baseline. The proposed architecture achieves a 14.5 dB SI-SNR improvement, only 0.7 dB below the baseline, while reducing model size by approximately nine-fold. Evaluation on commercial mobile CPUs demonstrates a 4.3-fold reduction in real-time factor, supporting practical on-device deployment. Compared with random search, RL-NAS achieves 92% higher search efficiency. Architectural analysis reveals that shallower but wider temporal convolutional blocks and aggressive visual pathway pruning can maintain enhancement performance while substantially reducing computational requirements. These findings demonstrate the potential of reinforcement learning-guided architecture search for developing lightweight, efficient and deployable multimodal speech enhancement systems.

Spatial semantic segmentation of sound scenes (S5) jointly performs audio source separation and tagging on a multichannel audio mixture. Therefore, an S5 system can be decomposed into an audio tagging model and a separation model. The baseline S5 system in task~4 of the DCASE 2025 challenge uses M2D and ResUNet models for tagging and separation, respectively. These are large in overall size and therefore not appropriate for use in embedded devices. The main aim of this paper is to assess the performance loss in an S5 system when using lightweight tagging and separation models instead. To this end, we first compare the tagging performance of lightweight audio tagging models, namely PANNs and DyMNs, in terms of label prediction accuracy against the baseline M2D model. Secondly, we assess separation performance in terms of the scale-invariant signal-to-distortion ratio of the baseline ResUNet separator relative to that of a convolutional-based separator using SuDoRMRF. Finally, we combine the tagging and separation models across various configurations and assess the overall performance of the S5 system using the class-aware signal-to-distortion ratio.

Speech remains intelligible to listeners under heavy temporal degradation, but whether speaker identity survives is largely unexplored. We evaluate six automatic speaker verification (ASV) models, trained on temporally intact speech, and tested under three degradations that remove different combinations of temporal cues. The models form three pairs, one per architecture family (convolutional, transformer, conformer), each pairing a waveform-input with a mel filterbank-input model. All models degrade, but robustness tracks input representation rather than pretraining scale or model size. In all nine family-by-condition comparisons the filterbank-input model degrades less, with eight of nine 95% confidence intervals excluding zero, even when it is smaller or has less pretraining than its waveform counterpart. In the controlled convolutional pair, where the models differ only in input representation, the filterbank-input model's equal error rate rises from 1.1% to 17.1% under the heaviest tested degradation, against 2.9% to29.7% for its waveform counterpart.

Generative creative interfaces routinely report the user's position within a space of possibilities. Such displays imply that the quantities shown were measured from the artefact; in a growing class of systems they are instead predicted by a language model describing its own output, and the two are indistinguishable on screen. Most creative domains lack an agreed measurement against which such a self-report can be checked; audio admits one. In a measurement study of a working LLM-driven timbre interface, what a user can cause proves separable from what the interface reports. The first is exactly characterisable: the reachable set is a zonotope whose dimension increases by one per sound placed, so the space the model may act within is demonstrably the user's own. The second is inaccurate. The displayed position differs from measurement by approximately three times the width of that region, and the error has two independent sources: model-predicted coordinates, whose accuracy varies twofold between the two models evaluated, and the interpolation step, which arises from combining synthesiser parameters and is unaffected by model quality. End to end, 26 of 36 refinement instructions produced no perceptible change, and under the weaker model the facility performed no action while reporting convergence. Effective agency (what a user can cause) and epistemic agency (whether they can know it) are therefore separable and independently engineered: structurally enforced constraints survive model failure, whereas guarantees resting on the model's own report do not. The failure mode that remains is silence presented as success. We close with a short protocol for building and auditing such displays.

Pretrained audio encoders are reused for downstream tasks that are often unknown when the encoder is trained, so their usefulness depends partly on which signal properties survive the pretext objective. We study this retained information through paired source reconstruction. Using a shared Stable Audio Open latent diffusion decoder, we reconstruct five-second, 44.1-kHz stereo music from frozen representations produced by supervised classifiers (VGGish, ConvNeXt), an audio-text contrastive model (CLAP), and a waveform-reconstruction model (EnCodec). These objectives impose different pressures to preserve source detail, while their exposed interfaces vary substantially in temporal and spectral resolution. Evaluating on the Million Song Dataset (MSD), we find clear differences in reconstructability across encoder families, while within encoder comparisons show improved recovery when finer temporal or spectral structure is exposed. Even compressed task oriented embeddings support reconstructions that preserve measurable source specificity and high level musical content.

Wearable augmented reality (AR) assistants are moving toward continuous real-world interaction, where they perceive the user's activity through first-person video and audio and provide timely spoken guidance without being explicitly asked. While proactive video assistants, spoken dialog systems, and egocentric task understanding have each advanced rapidly, existing systems do not address the joint problem of deciding when to speak and what to say from continuous first-person streams. We introduce EgoVoice, a framework for training and evaluating proactive egocentric spoken assistants. From HoloAssist video recordings of real human instructors, we construct clean audio streams through source separation and speech resynthesis, and convert each video session into a format where the model must decide at each moment whether to remain silent or provide spoken guidance. We fine-tune an omni-modal LLM with our data, and further improve its proactive intervention behavior with direct preference optimization. Experiments across closed and open-source models show that existing systems rarely produce well-timed, meaningful proactive interventions, while EgoVoice yields clear improvements in intervention timing, content relevance, and human preference over the zero-shot backbone.

Recent full-duplex spoken dialog models enable simultaneous listening and speaking, but fine-grained models still advance their backbone autoregressively at every interaction frame. We introduce DiffuPlex, a rolling masked diffusion framework that reduces this sequential computation by predicting multiple future user and assistant frames in a single backbone wake. DiffuPlex consumes only a confident prefix of each predicted future while interaction continues at the original frame rate. As user speech arrives, it checks the corresponding user predictions and, when the interaction diverges, preserves already played assistant content while revising only the unplayed future. We consider two inference policies over the same predictor: DiffuPlex-LISTEN consumes multiple future frames when they predict assistant silence, whereas DiffuPlex-SPEAK can also consume predicted assistant speech. Across full-duplex interaction and spoken-language evaluations, DiffuPlex substantially reduces sequential backbone computation while largely preserving interaction behavior and general capability. DiffuPlex-LISTEN and DiffuPlex-SPEAK achieve $1.46\times$ and $1.59\times$ deployment-path wall-clock speedups and $1.61\times$ and $1.80\times$ Core LM speedups, with all measured backbone invocations completing within the 80ms interaction interval. Human evaluation shows that LISTEN preserves speech naturalness and conversational quality, while SPEAK retains conversational quality with some degradation in speech naturalness.

Music understanding often requires comparing excerpts and reasoning about relationships among songs, sections, and stems. However, existing large audio-language models (LALMs) and music question-answering datasets typically operate on single recordings or compare independently sampled tracks with no known production relationship. We introduce STEMMA, a multi-audio music question-answering framework built around production provenance: whether excerpts originate from the same track or section, and which stems belong to which mixtures. Because such relations are sparse under conventional audio-first sampling, STEMMA adopts a relation-first construction strategy: it first specifies a target relation and then queries the catalog for excerpts that satisfy it and hard negatives that do not. Labels are determined directly from catalog provenance rather than generated by a language model from metadata. We build STEMMA-Bench for evaluation and a track-disjoint training set, STEMMA-Instruct. Fine-tuning two LALMs on STEMMA-Instruct improves multi-audio reasoning, with the largest gains on structural relations directly determined by the catalog, while preserving single-audio music understanding.

Open-Vocabulary Audio-Visual Event Localization (OV-AVEL) labels each video segment with an event class, including classes that were never seen during training. The dominant pipeline uses a frozen multimodal foundation model (e.g. ImageBind) to embed the visual frame, the audio mel-spectrogram, and each candidate class name into a shared space, then computes two cosine similarities for each segment against each class: visual-text and audio-text. Existing methods then collapse this pair into a single scalar score with a fixed rule (geometric mean, weighted average) before taking the argmax. Instead, we compute complex-valued similarities and learn their fusion using a complex-valued neural network (CVNN). Each modality's standard representation becomes the real part of our pipeline, and a paired companion stream supplies the imaginary part. We use imaginary part of iHSV for visual modality and CycleGAN-translated phase spectrogram for audio modality as these companion streams. This results in two complex similarities, which are then fused. While the vision and audio encoders remain frozen, only the temporal-attention blocks and the fusion CVNN are trained. The four-stream complex architecture sets a new state of the art on both OV-AVEL benchmarks. On the open (unseen-class) split of OV-AVEBench we reach 66.5/59.1/54.1% Acc/Seg-F1/Event-F1 (+1.6/+4.1/+6.6 over the previously reported fine-tuned baseline), with consistent gains for seen classes as well. We also modify AVE dataset for this task and observe that our architecture reaches 60.7/51.9/50.4% Acc/Seg-F1/Event-F1, achieving state-of-the-art OV-AVEL results on it as well. We also propose a two-stream alternative, which also sees great improvements over the baseline.

Detecting intra-operative speech impairment during awake craniotomy is essential for preserving language function. However, automated detection remains challenging because operating-room recordings contain substantial acoustic interference, clinically relevant speech events are rare, and available cohorts are small and heterogeneous across speakers. This study presents a systematic component-wise evaluation of a pipeline for distinguishing dysarthric from no-trouble speech in the DATABRASE corpus of awake-craniotomy recordings. The pipeline incorporates speaker diarization to isolate patient speech, a multi-view representation combining handcrafted acoustic descriptors with multilayer wav2vec 2.0 embeddings, speaker-conditional normalization and transferability-based feature selection to improve cross-speaker robustness, and a cascaded classifier comprising a gradient-boosted first stage and a neural second stage. Evaluation was conducted under strict speaker-independent conditions using leave-one-speaker-out cross-validation. The results show that cross-speaker performance is influenced more strongly by the speech representation than by classifier choice. The AUCs of three classifiers differed by no more than 4.7%, whereas replacing conventional acoustic descriptors with the multilayer self-supervised representation produced AUC improvements of 18.2%-26.1%. Diarization-conditioned feature extraction and the proposed classifier cascade provided additional consistent gains. These findings indicate that reliable patient-specific speech isolation and strong pretrained representations are more important than increased classifier complexity in low-resource intra-operative settings. They also quantify the potential performance gains that may be achieved through patient-specific preoperative calibration.

Existing singing pitch correction approaches rely on target melodies or accompaniment tracks, which may be unavailable in practice. We formulate reference-free singing pitch correction as a music-constrained sequence editing task that determines whether and how each note should be corrected from the input performance alone. A pretrained symbolic music encoder with lightweight singing-domain adapters produces contextual representations of the singing MIDI. Based on these representations, two lightweight correction heads jointly model correction necessity and signed pitch modification through a factorized pitch-editing distribution. An input-dependent tonal prior derived from the estimated key distribution then reranks the candidate offsets, favoring tonally compatible corrections without target melodies or ground-truth key annotations. Experiments on real paired amateur and professional singing recordings show that the method improves note-level pitch accuracy from 73.25 to 83.43, outperforming a context-based baseline by 3.99 percentage points while balancing error correction and preservation of correctly performed notes.

Natural-language style descriptions provide an interpretable interface between large language models (LLMs) and controllable text-to-speech (TTS). However, using descriptions as pseudo-labels compresses target acoustics into text, and descriptive fidelity need not imply effective control of a particular synthesizer. We empirically show that speech-text alignment only weakly predicts downstream acoustic similarity among candidate instructions for the same utterance. We therefore propose Speech-Rewarded Style Planning (SRSP), which trains a text-based style planner through a frozen downstream TTS model. Given dialogue history and response text, the planner generates candidate instructions and is optimized with group-relative policy optimization (GRPO), using the teacher-forced likelihood of target speech tokens as the reward. On an English subset of the ISCSLP 2026 CoT-TTS corpus, SRSP achieves higher speech-style and emotion similarity to target speech and lower mel-cepstral distortion than the Base LLM and target-audio-informed captioning baselines. LLM-based expressive speech evaluation further shows gains over all baselines in contextual appropriateness and reference consistency.