Papers

Residual scalar quantization can produce highly dependent indices when successive stages reuse identical, unscaled grids. This case study examines that behavior in a frozen neural audio codec and develops an exact recoding procedure that reduces storage while preserving its decoder inputs. Analysis of the repeated-grid recurrence identifies predictable stage sequences; finite-precision checks motivate preserving every original index through explicit exception records. The model weights and reconstruction graph remain fixed. On the studied 62- and 512-file cohorts, a predictor, adaptive coding and explicit file selection reduce complete stored bytes. The final 512-file rate is 129.591 kbps including headers, 25.943% below an original-index XZ control. Original recorded CPU runs establish bitwise latent and waveform preservation; independent readers verify exact index and metadata recovery in subsequent saved-stream experiments. Later coding choices use previously studied cohorts, and fresh encoding requires generating three alternatives. The engineering result is implementation-specific storage reduction with exact index recovery; the underlying lossy reconstruction quality remains unchanged.

Generating datasets with audio plugins requires preserving how each audio observation was produced. vst-bench-ml is a Python package that samples plugin parameters, coordinates rendering through a separately installed host, and records software identities, rendering settings, and output hashes for later verification. A sealed manifest binds this evidence and supports offline integrity checks without loading the plugin. A companion repository provides a sample-free VST3 test plugin, a deterministic offline host, and additional checks linking rendering attempts to recorded outputs. On a pinned macOS arm64 and Python 3.12 stack, functional validation of the pre-release versions reproduced package and host tests. Two runs in fresh processes reproduced the same render-plan hash and all eight expected output hashes, covering four sampled parameter configurations and two fixed MIDI signals; 12 negative cases also passed. Later licensed 0.1.0 release candidates passed producer-run continuous integration and local validation. Version 0.1.1 adds public Git source repositories, package-index installation, and user documentation; a separate release check reproduced the expected plan and output hashes. These checks establish integrity and repeatability within the tested setup; they do not establish portability, scientific suitability, or perceptual quality.

Lumped-element networks give transparent, passive models of acoustic impedances whose low- and high-frequency limits can be read from the circuit, but a systematic, arbitrarily accurate construction for non-rational impedances is lacking. Continued fractions, one- and two-point Padé approximants and their identity with ladder networks and Cauer synthesis are introduced in tutorial form and applied to two examples: the radiation impedance of a baffled circular piston and the throat impedance of a horn with a piston mouth. The even contraction of the piston's continued fraction has positive coefficients, certified by interval arithmetic for the first 100 pairs, and yields a ladder of positive resistances, masses and compliances; with 14 elements the error is 4e-4 up to ka = 5. Beranek's four-element network is, to about 2%, the order-two two-point Padé approximant. A horn modelled as a cascade of waveguide sections is a continued fraction whose tail is the mouth impedance; the infinite exponential horn is a periodic continued fraction, an exact high-pass ladder. The piston networks are passive at every certified order (up to 202 elements) and are extended without changing existing elements; for a finite horn the error of the mouth network passes one to one to the throat impedance.

Speech-capable language models inherit most of their knowledge from text-rich language backbones, and continued model maintenance is likewise dominated by text updates. Knowledge acquired through a text update can nevertheless remain difficult to access when the same query is spoken. We formulate this mismatch as post-update speech access repair, where the knowledge-bearing checkpoint stays frozen and a detachable operator repairs speech-only access in one model pass. We introduce Affine Fields for Interface Repair across Modalities (AFFIRM), which first learns and freezes a base correction for speech hidden states, then adds query-specific corrections with zero mean over the operator-training queries. Paired speech and text states provide the coordinates that condition these corrections on each query. On balanced synthetic-speech PopQA, paired comparisons do not establish a clear advantage for AFFIRM over global SALSA steering or low-rank ReFT interventions under the same frozen single-state constraint. On recorded SLUE-SQA-5 questions with text passages, a full-scale PopQA-fitted AFFIRM operator transfers without refitting and improves ROUGE-L by 0.0941 over its paired No Repair baseline. Low-rank ReFT interventions also transfer positively, whereas the matched SALSA variant falls 0.0707 below its paired baseline. These results show that text acquisition and speech access can diverge after an update and should be verified and repaired separately.

Modern audio watermarks achieve strong robustness with learned encoder-decoder models, but their behavior under new processing is established empirically. CochlearMark is introduced as a training-free audio watermark that uses auditory-nerve phase-locking as its feature representation. A small waveform perturbation is optimized against a keyed random projection of this representation, and the same projection is used for detection, with no trained encoder or decoder. Because phase-locking represents temporal fine structure, the resulting robustness has a direct physiological interpretation: the mark is retained when this structure is preserved and lost when it is disrupted or rebuilt. Across speech, music, and environmental audio, CochlearMark achieves signal-edit robustness and imperceptibility comparable to trained neural baselines while requiring only a lightweight computational footprint. Among the evaluated training-free methods, CochlearMark is the only one to retain substantial detectability under neural-codec re-encoding.

Automatic prompt optimization (APO) has shown strong effectiveness for text-based LLMs and LLM systems, but its application to audio-visual understanding systems requires suitable textual interfaces through which multimodal reasoning can be exposed to text-based optimizers. We investigate how the design of textual interfaces affects prompt optimization for multimodal reasoning, on the audio-visual question answering task. We compare three interface designs, including a direct inference system consisting of a single omni-modal large language model (OmniLLM), a captioner-reasoner system, and an agentic architecture in which a primary agent orchestrates subagent OmniLLMs. Experiments with Qwen3-Omni and Gemini show that direct prompt optimization provides model-dependent gains; the captioner-reasoner system and the agentic system expose more interpretable optimization interfaces but are harder to optimize, where error analysis of cases affected by optimization is conducted. Our study highlights textual interface design as an important factor in making prompt optimization effective for omni-modal reasoning systems. Our implementation details are specified in GitHub repo.

Reliable fault detection is crucial for industrial machinery in large-scale facilities, where unexpected failures can cause costly downtime and production interruptions. Although acoustic monitoring provides a scalable solution for machine condition monitoring, its deployment is often constrained by high false-positive rates. We propose a two-step approach to automatic anomalous sound detection: first, when only normal operating audio samples are available, a regularized unsupervised model is deployed; then, after the first machine fault, a few-shot adaptation strategy exploits the anomalous samples acquired during the fault to improve the system. This adaptation increases the area under the precision-recall curve (AUPR) from 23.0 to 64.0 and improves detection in the low false-positive-rate region, with the partial area under the receiver operating characteristic curve (pAUROC) increasing from 39.6 to 72.8. We also show that covariance regularization improves both prediction accuracy and training stability in the unsupervised model. These results provide a practical strategy for acoustic anomaly detection systems, enabling improved accuracy and reduced false alarms.

Lipreading remains challenging due to the inherent ambiguity of visual speech cues and their sensitivity to head pose and video resolution. In this work, we propose FLARE, a framework for robust visual speech representation learning via cross-modal distillation and joint face-lip modeling. FLARE leverages semantically rich Whisper representations as the acoustic teacher signal, converting them into discrete distillation targets through a random-projection quantizer and pretraining the model with a masked prediction objective. To capture complementary visual information, FLARE employs a dual-stream architecture that jointly processes lip region-of-interest (ROI) and full-face inputs, where the lip stream captures finegrained articulatory motion and the face stream encodes broader facial dynamics. During pretraining, auxiliary audio input is incorporated to facilitate cross-modal alignment, and modality dropout is applied to encourage robust visual-only representations. For downstream lipreading, FLARE serves as a visual encoder coupled with either a lightweight Transformer decoder or an LLM-based decoder. Evaluated on LRS2, LRS3, and the out-of-domain WildVSR benchmark, FLARE achieves strong lipreading performance, with word error rates of 11.7%, 13.2%, and 34.3%, respectively. A progressive design study confirms the contribution of each major component. Further analyses demonstrate that joint face-lip modeling improves robustness under degraded video resolution and large head pose variation, while yielding more phoneme-discriminative visual representations. Code and pretrained models are available at https://github.com/jxzhanggg/FLARE.

Text-to-speech (TTS) based on masked diffusion models (MDMs) achieves competitive zero-shot speech synthesis performance. Unlike autoregressive (AR) models, MDMs define a corruption process that maps clean speech-token sequences to masked maps and learn a reverse denoising process to recover the reference tokens. However, MDMs commit to intermediate predictions that cannot be revised in subsequent inference steps. Consequently, prediction errors accumulate and propagate, degrading synthesized speech quality. We present Revise-TTS, a single-stage framework that applies hierarchical binary quantization (HBQ) and generative refinement networks (GRN) to zero-shot speech generation. First, we replace finite scalar quantization (FSQ) with HBQ to introduce a scalable binary representation for speech tokens with an explicit coarse-to-fine structure. Second, GRN is trained to reconstruct the token sequence from a mixture of ground-truth and random tokens. During inference, iterative refinement enables self-correction by randomly discarding a small proportion of previously predicted tokens with new predictions. Experiments and ablation studies demonstrate that Revise-TTS achieves competitive zero-shot speech synthesis performance using HBQ. It can further be extended to any MDM-based TTS task as an alternative training and inference method.

Post-training often improves the performance of large audio-language models on audio understanding and reasoning benchmarks. However, these improvements may reflect adaptation to specific question formats, such as multiple-choice questions, rather than genuine advances in audio understanding and reasoning capabilities. To investigate what LALMs actually learn from post-training, we conduct controlled experiments that disentangle improvements in audio understanding from adaptation to specific question formats. Across different models and post-training methods, we find that much of the reported improvement persists even without informative audio, but largely disappears when the question format changes. These findings suggest that post-training often teaches models to exploit task-specific textual cues and answer patterns rather than developing generalizable audio understanding and reasoning capabilities.

Neural instrument synthesizers can clone the timbre of a solo instrument, letting music producers make music with the sound they desire. However, an instrument of interest is often embedded in a fully produced, multi-instrument recording, beyond the reach of existing methods. Source separation before timbre cloning is an approach, but it may suffer from separation artifacts and is also likely a detour for the task. In this work, we propose Point-and-Play, a system that clones the timbre of a cued source directly from a polyphonic mixture and uses it to synthesize new pitches and velocities. To achieve this, we introduce a Cue-Gated Statistics Pooling (CGSP) module that extracts the timbre code of the cued instrument from adapted CLAP embeddings of the music mixture. This extracted timbre code is then used to condition a separately trained note synthesizer together with pitch and velocity inputs. Experiments show that our proposed system outperforms separate-then-clone baselines in terms of timbre fidelity and synthesis quality, achieving comparable performance with that cloned from a clean solo instrument.

Multimodal Emotion Recognition (MER) aims to understand human emotions by integrating information from multiple modalities such as speech and text. Psychological studies suggest that emotions exhibit latent hierarchical structured relationships in their semantic representations. However, most MER approaches rely on Euclidean representations, which capture fine-grained cross-modal interactions but do not explicitly provide a geometric prior for organizing hierarchical emotional structures. To mitigate this gap, we propose GeoMER, a joint Euclidean-Hyperbolic modeling framework that introduces hyperbolic emotion-core routing to capture higher-level, abstract emotional representations for MER. The proposed framework jointly models token-level cross-modal semantic interactions and latent emotion representations in two complementary geometric spaces. Specifically, GeoMER consists of a Euclidean Multimodal Fusion Module (EuclFusion) for multimodal representation fusion and a Hyperbolic Geometric Core Routing (HypGeoRoute) module that performs geometry-aware core assignment and aggregation in a hyperbolic manifold to capture more abstract feature representations. The HypGeoRoute module complements the Euclidean token-level modeling. We experiment on the IEMOCAP multimodal MER benchmark and demonstrate that the proposed method achieves the best performance, with 83.54% weighted accuracy and 83.35% unweighted accuracy, outperforming Euclidean baselines, including MemoCMT.

We investigate how speech token representations can support both understanding and generation within a unified speech language model. We present PhonoVoxtLM, built on the existing VoxtLM framework using phonological tokens designed to retain linguistic content and prosodic information. The existing tokenizer is jointly trained for speech recognition and resynthesis. Its single-codebook units serve as both speech inputs and generation targets, without additional speech-token prediction modules. Experiments show improvements over HuBERT-based baselines across understanding and generation under different training-data scales. Further comparisons with WavLM-based tokens reveal trade-offs between recognition and synthesis, highlighting the importance of evaluating shared representations across both capabilities. We further extend the framework to spoken question answering through fine-tuning with replay. The adapted model improves answering performance over similarly adapted HuBERT baselines while retaining advantages on the original speech tasks. Our findings support single-codebook representations designed for both content and prosody as a practical choice for unified multitask speech modeling.

Deep learning models of speech seldom represent the source-filter process by which speech is produced, relying instead on latent features with no physical correspondence to articulation or phonation. We propose SoFilNet, a self-supervised, differentiable Source-Filter Network that estimates the parameters of speech production directly from the raw waveform: a source, specified by fundamental frequency (F0), voicing and the aperiodicity of five frequency bands, and a filter, specified by ordered formant poles, additional poles that shape the rest of the spectral envelope, and antiresonance zeros. The model is trained only to reconstruct the input speech; no formant or pitch annotation is used. Since per-bin spectral losses favor blurred spectra wherever the model is uncertain, we train F0 and voicing against autocorrelation pseudo-labels refined with the model's own filter, aperiodicity against the periodic fraction of each band, and formant sharpness against spectral contrast. On VTR-FORMANTS, the estimated formants attain the lowest mean error among the reported methods, including supervised trackers, and on the unseen PTDB-TUG corpus the F0 frame error matches the best classical tracker and is lower than that of all learned trackers. The parameters are also consistent with speech production theory: aperiodicity rises with frequency and orders phone classes by the turbulence of their excitation, and the zeros deepen on nasal consonants, the lower one rising with place of articulation.

Can an integrated optical readout improve the useful operating range of a miniature microphone? We address this question with an analytical model of a diaphragm carrying the sensing arm of a Mach-Zehnder interferometer (MZI). The model links pressure-induced bending to waveguide elongation and then to photocurrent, allowing mechanical sensitivity and optical detection limits to be assessed separately. Optical shot noise and back-cavity thermomechanical noise set the pressure resolution, while a 10% total harmonic distortion criterion defines acoustic overload. Two operating regimes are examined: a compact MEMS microphone with 1 mW optical input and a measurement microphone with 100 mW optical input. For the compact case, a 1 mm diameter, 1 micrometer thick, stress-free silicon diaphragm and 100% modulation efficiency give a predicted signal-to-noise ratio of 78 dB and acoustic overload pressure of 124 dB. The higher-power case lowers the modeled noise floor and increases dynamic range relative to the selected measurement microphone benchmark. These estimates rely on favorable assumptions and do not establish a substantial overall advantage in conventional microphone applications. Remote optical interrogation offers a separate motivation for harsh-environment sensing, but the temperature tolerance of the proposed sensing head remains unverified. The analysis identifies the design tradeoffs and measurements needed to test the predicted performance.

We present a perceptual audio coder in which the entropy-coding model and the bit allocation co-adapt across blocks without model side information. Mantissas are coded by a hybrid of bit packing and arithmetic coding with per-band probability models learned online. Encoder and decoder accumulate identical histograms from decoded blocks and advance the same model state. The model enters bit allocation directly: per-band rate–distortion frontiers combine measured perceptual distortion with entropy-coded rate estimates under the current model, and a dynamic program attains the exact minimum of total objective perceptual distortion for each block's budget under the model's rate estimates, with reservoir carry-over. Under a strict per-block rate, entropy-coding-aware bit allocation lowers this distortion (band noise-to-mask ratio in linear scale) by 15-24% relative to the same solver without it, and co-adaptation removes a further 3-4% (up to 15% on oboe) that a backward-adaptive coder alone cannot realize in one pass; a MUSHRA test at 80 kbps per channel confirms the ranking.

Active noise cancellation (ANC) is common in headphones but rare in hearing aids, and its value for understanding speech in noise is not established. Algorithms that improve signal-to-noise ratio (SNR) act only on the hearing aid's amplified path, while unprocessed sound also reaches the eardrum directly through the vent, limiting the SNR improvement that is realized. Our earlier simulations predicted that attenuating this direct path with ANC would substantially increase the benefit of such processing. We tested that prediction in three hardware conditions: a standard receiver-in-canal (RIC) hearing aid, and an ANC-capable RIC with ANC off and on. Each was fit to NAL-NL2 targets for an N3 hearing loss and recorded on KEMAR with an AI speech-enhancement algorithm (AI-SE) off and on, in seven realistic scenes above 70 dB SPL. Speech and noise spectra were estimated from each recording with a binary time-frequency mask derived from the clean speech. AI-SE produced a similar peak benefit of about 4 dB in the high frequencies in all three conditions, but that benefit stayed at or above 1 dB only above 1500, 600, and 130 Hz for the standard RIC, ANC RIC (ANC off), and ANC RIC (ANC on), respectively, closely tracking the measured attenuation of the direct path. With AI-SE on, ANC RIC (ANC on) improved SII-weighted SNR by 5.1 dB relative to the open ear, compared with 3.3 dB for ANC RIC (ANC off) and 1.6 dB for the standard RIC. With AI-SE off, enabling ANC added only 0.2 dB. ANC is therefore an enabling technology: on its own it added little, but by attenuating the direct path it allowed SNR-enhancing processing to deliver substantially greater benefit at the eardrum, extending that benefit to much lower frequencies.

Continuous latent representations underlie recent diffusion-based lyrics-to-song and text-to-music systems, yet aligning lyric conditions with these latents remains difficult: existing systems resort to generator-side machinery such as sentence-level timestamps, auxiliary alignment objectives, or semi-autoregressive decoding, while the latents they predict are still produced by autoencoders trained purely for reconstruction. We address this difficulty from the representation side. We propose MSVAE (Music Semantic Variational Autoencoder), which structures its continuous latent around two complementary semantic streams: work-level compositional content and lyric-level linguistic-vocal content. Partitioned dual-stream alignment makes both streams explicit by aligning disjoint latent subspaces with frozen lead-sheet and ASR references, while a semantic-first gated residual posterior derives its base from a frozen music-SSL encoder and admits acoustic information only through gated, bounded corrections. MSVAE achieves competitive stereo reconstruction and stronger MARBLE probing than continuous-autoencoder baselines, and improves downstream lyrics-to-song and text-to-music generation. Notably, with only an untimed phoneme prefix and no lyric timing, MSVAE yields the lowest phoneme error rate, whereas all baseline representations require sentence-level alignment to produce intelligible vocals.

Dynamic spatial audio reasoning requires a model to associate changing acoustic observations with the correct source and recover its spatial state at a queried time or interval, a challenge intensified by overlapping speech and source motion. To study this problem, we introduce Sonic-Bench, comprising 2.8K first-order Ambisonics clips and 28.8K target-specific question-answer pairs across 480 simulated indoor scenes. The benchmark covers localization, distance estimation, and motion reasoning in single- and dual-source settings. We further propose SonicWorld, which integrates physical, learned acoustic, and semantic features into persistent source states through learned transitions and recurrent updates, and projects them into an Omni LLM for spatial reasoning. Compared with Spatial-Omni on the Sonic-Bench test split, SonicWorld improves direction accuracy from 31.98% to 34.69% and reduces azimuth error from 57.22{\deg} to 48.01{\deg}. In zero-shot transfer to Spatial LibriSpeech, SonicWorld retains strong localization performance, achieving 73.98% direction accuracy and 19.29{\deg} azimuth error, compared with 27.14% and 70.74{\deg} for Spatial-Omni.

Full-duplex spoken language models (FD-SLMs) support simultaneous listening and speaking, but often lag behind half-duplex spoken language models (HD-SLMs) in general capabilities. We propose Dialogue-to-Duplex (D2), a framework that converts pretrained HD-SLMs into high-capability FD-SLMs at fine temporal granularity. D2 adapts the original audio encoder to streaming through decoder-free representation distillation and learns full-duplex interaction through supervised fine-tuning. We systematically study the effects of chunk size on full-duplex interaction, general capabilities, and inference computation. We apply D2 to two distinct HD-SLM architectures, Qwen3-Omni and LLaMA-Omni 2. At a 12.5Hz update rate, Qwen3-Omni-D2 achieves 99.2% turn-taking success with 521-ms latency, 99.5% interruption success, and a state-of-the-art VoiceBench score of 70.1.

Effective vocal coaching connects performance issues with corrective advice, but recordings paired with expert diagnoses and advice are scarce. Online singing communities provide a potential source of supervision, but observations and advice are scattered across unstructured comments. We introduce PitchIn, audio-language models for natural-language vocal coaching, and the PeerCoach Dataset: 14,445 instruction examples covering Focused QA, Overall Feedback, and Recognition QA. We compare target combinations on Qwen2.5-Omni and train PitchIn on Focused QA and Overall Feedback using Qwen2.5-Omni and MOSS-Music. On VocalCoachBench, the proportion of requested diagnosis-correction pairs matching an expert issue with valid advice rises from 21.3% to 61.0% for Qwen2.5-Omni and from 31.4% to 51.2% for MOSS-Music. Direct training on source comments yields only 19.8% for Qwen2.5-Omni. For Qwen, Focused QA alone reaches 65.6%, but adding Overall Feedback improves alignment with expert reviews and human preference. Both models improve expert-review alignment over pretrained baselines and are significantly preferred for usefulness (92.0% and 67.3%, respectively). Demo examples are available at https://hayeonbang.github.io/PitchIn.

Speech enhancement with wearable and edge devices requires architectures that balance enhancement quality against power and memory resources. Spiking Neural Networks (SNNs) offer a promising path towards this goal through sparse, event-driven computation, but existing SNN-based speech enhancement methods are largely restricted to single-channel processing and therefore cannot exploit the spatial diversity available from microphone arrays. In this work, we propose a fully SNN framework for multichannel speech enhancement based on a fully connected architecture (FC-SNN) with sequential spike encoding, derived from a compact artificial neural network (ANN) baseline via leaky integrate-and-fire neurons. We characterize how enhancement performance scales with model size and spike-train length relative to ANN and classical beamforming baselines, and report the resulting parameter counts and computational complexity. We further compare direct and rate spike encoding, revealing a clear trade-off between event-driven computation and enhancement performance, and use these results to identify favorable configurations for edge devices.

Speech timing is fundamental for speech models to recognize words accurately, but how word-timing information is represented in speech models remains underexplored. We study this question using the framework of representational geometry, which investigates whether and how concepts are represented in low-dimensional subspaces of the representation space. Prior work in cognitive neuroscience has shown that when HuBERT feature sequences are averaged across words of the same duration, they form near-circular trajectories from word onset to word offset in a two-dimensional "temporal subspace." Here we show that this circular structure of word timing is a general property of deep speech representations, by demonstrating its emergence in nine additional speech models covering both supervised and self-supervised. We then move beyond averaged word segments and show that in HuBERT, the circular trajectories are consistent enough at the level of individual words to perform word segmentation. Our approach, Rotary Segmentation (RoSeg), achieves 64.3% boundary F-score on Librispeech by detecting word boundaries when cycles complete in the temporal subspace. Although supervision is needed to identify a good subspace, nevertheless this result demonstrates much stronger implicit word knowledge in HuBERT than has previously been shown.

Temporal audio grounding requires locating every occurrence of a queried target, making both occurrence counts and interval boundaries essential. We propose RefTAG, a post-training framework for verifiable temporal self-correction. Its Reflection Chain-of-Thought (Re-CoT) protocol combines draft intervals, discrete count and boundary actions, and revised intervals in one response. Supervised fine-tuning on these trajectories (Re-CoT-SFT) teaches the protocol; reinforcement learning (RL) jointly rewards final localization and revision outcomes. Across 8,228 queries, Re-CoT-SFT improves query-weighted mIoU over final-answer SFT by 1.74 percentage points. RL further improves TAG-Bench multi-target mIoU by 0.40 points and R@0.5 by 0.86 points. The final model achieves 38.50% TAG-Bench mIoU and 51.01% mean mIoU across three datasets, outperforming the evaluated open-source MOSS-Audio and FireRedAudio models and the commercial Gemini-3.1-Pro on both measures. These results demonstrate that explicitly coupling count- and boundary-aware revision with joint reward optimization provides an effective and verifiable path to temporal audio grounding.

Spatially encoded speech conveys not only linguistic information but also spatial cues. This enables listeners to distinguish and localize speakers, contributing to more natural and realistic auditory experiences in applications such as remote conferencing and immersive media. Neural speech codecs (NSCs) enable high-quality reconstruction of monaural speech at increasingly low bitrates while preserving linguistic information, but their ability to effectively preserve spatial cues remains limited. We propose FOAdapter, a plug-and-play spatial tokenizer that leverages a pretrained monaural speech codec to efficiently represent both linguistic and spatial information in first-order ambisonic (FOA) speech at low bitrates. Comprehensive evaluation demonstrates that FOAdapter consistently preserves spatial information across all evaluated codecs. Source code and audio examples are available on our demo website.

Automatic Speech Recognition (ASR) systems are commonly designed to transcribe linguistic content independently of speaker identity. However, speaker-specific information can persist within the latent representations of ASR models, raising privacy concerns when training data includes sensitive individuals. While machine unlearning has been studied in several speech tasks, applying speaker unlearning to ASR remains challenging due to the strong entanglement of linguistic and speaker-related information in speech representations. In this work, we conduct a layer-wise analysis of Self-Supervised Learning (SSL) based ASR representations, revealing that speaker identity is primarily encoded in intermediate encoder layers. Based on this observation, we explore a feature-level unlearning strategy that targets these identity structures. Experiments on the VCTK corpus show that the proposed approach removes the influence of specific speakers while maintaining robust transcription performance.

Linking speakers across independently recorded rooms is difficult because microphone, room, and codec mismatch distort speaker embeddings. We study visitor linking without visitor enrollment or identity labels in simulated outpatient workflows. The proposed method combines per-device embedding centering, concurrency constraints, room-type-specific visit constraints, and a clustering threshold calibrated from concurrent visitors and identified staff voices. On simulated days built from VOiCES recordings, with companions, return visits, walk-ins, silent visits, and acoustic leakage, it improves B$^3$-F1 over the best label-free baseline for all 11 extractors, by 5.3-14.6 points (all 95% intervals exclude zero). ECAPA-TDNN reaches 86.6, compared with 76.4 for the best label-free baseline and 87.7 for the best oracle-informed baseline. Gains persist across the reported workflow perturbations, but depend on removing duplicated cross-room speech: without this step, the method wins for only 2 of 11 extractors. Single-visit patients also remain vulnerable to false merges. These results support structure-aware linking while identifying preprocessing and identity-safety limitations that must be addressed before clinical deployment.

We consider the problem of mono-to-stereo audio upmixing: generating a high-fidelity two-channel stereo signal from a single-channel recording. The task is fundamentally ill-posed i.e., many stereo signals are consistent with a given mono recording. An effective upmixer must therefore generate a realistic, musically coherent stereo image while preserving the underlying audio content. Previous approaches have relied on signal-processing driven decorrelation, ambience extraction and supervised learning, offering limited spatial control or requiring task-specific training. In response, we propose Aura (Audio Upmixing using Reward-tilted Flows), which reformulates upmixing as a generative inverse problem using a pretrained flow prior. To enable controllable spatial imaging, we design a reward-tilted posterior sampler whose reward penalizes divergences between generated and target distributions of inter-channel cues. Composing forward operators further enables joint restoration and upmixing without retraining or fine-tuning. Objective and subjective evaluations on real-world music recordings from MUSDB18-HQ demonstrate our method's effectiveness relative to several baselines.

Generating a full song from lyrics and a style description requires long-range musical planning without sacrificing vocal and instrumental detail. Low-rate discrete semantic tokens make this tractable, but quantization constrains the information available to acoustic rendering, while separately optimized stages prevent downstream generation objectives from shaping the planner. We propose joint discrete and continuous modeling for high-fidelity song generation. A VQ tokenizer supplies tokens for autoregressive planning, and a separately trained SSL-derived variational encoder supplies continuous semantic targets. Conditioned on the completed token plan and differentiable readouts of the planner's predictive states, a full-sequence flow model generates semantic features for acoustic rendering. Its flow-matching gradients reach the language model through the readouts, without differentiating through token sampling; the standard pathway keeps continuous features out of autoregressive feedback. On 192 shared prompts, its 48D configuration improves SongEval from 4.152 to 4.269 and reduces micro-PER from 20.26% to 14.47% relative to a discrete baseline, while receiving substantially higher expert listening scores. We further analyze the system through downstream semantic probes and conditioned reconstruction.

Music-understanding benchmarks should identify both the evidence available to a model and the operations a correct answer requires. We introduce HMU-Bench, a bilingual benchmark with notation and audio components. ABC Q&A contains 1,029 Chinese-English question pairs across six musical domains, labeled as extraction, single-step inference, or multi-step inference. Audio Q&A contains 460 question pairs on 92 recordings, spanning five scope-difficulty configurations. All music is Suno-generated, and questions and references are human-reviewed. Reference-based evaluation of seven omni models yields high notation scores for the strongest systems, but no model averages above 1.4 on the 0-4 audio scale. Notation and audio rankings show little correlation. A transposition case study also reveals octave errors missed by LLM grading. These findings motivate separate modality reporting and event-level checks for symbolic outputs. Data, annotations, evaluator prompts, and graded responses will be released at https://github.com/stepaudiollm/hmu-bench.