Audio-Language Models

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.

Large audio-language models (LALMs) exploit multimodal evidence, yet task-irrelevant audio can alter text-reasoning decisions when listening is unnecessary. Aggregate Accuracy can hide this paired drift because audio-induced repairs and damages may cancel. Paired drift analysis and targeted interventions identify architecture-specific, intervention-sensitive late audio pathways as actionable control points. We introduce ICAP-Gate, which applies mechanism-guided, task-conditioned control to each model's pathway. Across four LALMs, two reasoning benchmarks, and environmental-sound and natural-speech interference, ICAP-Gate has lower point estimates for Influence Rate and Answer Flip than ungated inference in all 16 full-split model--condition evaluations. Fixed suppression degrades automatic speech recognition (ASR) across all four models, whereas ICAP-Gate matches ungated ASR performance by preserving the pathway for explicit audio-demand instructions. ICAP-Gate has lower paired-drift point estimates than mitigation prompting in all four evaluated settings and provides competitive stabilization relative to eight-sample Self-Consistency while using one generation per query; in controlled ARC measurements, Self-Consistency incurs $7.0$--$9.2\times$ ungated latency. These results establish selective modality influence control as a design principle for robust multimodal reasoning.

Large audio-language models (LALMs) have achieved strong performance in general audio understanding, yet most are designed for monaural input and discard the inter-channel cues essential for spatial perception. In contrast, existing spatial audio-language models are purpose-built for spatial tasks and fail to capitalize on the general understanding capabilities of monaural LALMs. We present MiDashengLM-Spatial, the first open-source end-to-end unified audio-language model, to our knowledge, which supports both general audio understanding and spatial awareness within a single architecture. It extends MiDashengLM with a spatial audio encoder, Spatial-Dasheng, integrated through a hierarchical semantic-to-spatial conditioning module that injects intermediate semantic representations into the spatial branch at multiple depths while preserving the original semantic pathway. To provide spatial audio-language supervision at scale, we develop a data synthesis pipeline that renders diverse spatial acoustic scenes with scene-level spatial descriptions and question-answer pairs. Experiments show that Spatial-Dasheng achieves strong performance on sound event localization and detection in real-world scenes, and that MiDashengLM-Spatial substantially outperforms existing LALMs on spatial understanding and reasoning benchmarks. Meanwhile, it remains competitive with state-of-the-art 8B-scale LALMs on diverse monaural benchmarks, demonstrating that spatial awareness can be acquired without compromising general audio understanding. The source code and model checkpoint are available at https://github.com/xiaomi-research/midashenglm-spatial and https://huggingface.co/mispeech/midashenglm-spatial.

Large audio-language models (LALMs) fuse an audio encoder into a large language model (LLM) through multi-stage training. This coupling means that a new domain or a stronger LLM requires retraining, and their answers cannot be traced to what the model heard: a chain-of-thought is a post-hoc account. We propose Listen-to-Reason (L2R), an interpretable-by-design pipeline that passes audio to the LLM through an explicit, human-readable tree: small heads on frozen expert encoders map each chunk of a clip to semantically meaningful nodes on the tree (for speech, music and environmental sound), and a frozen text-only LLM answers from these nodes and an ASR transcript without hearing the clip. Every answer can therefore be traced to the nodes and transcript it read, and the nodes are causal: replacing the deciding node with a distractor overturns 78% of correct answers on SAKURA. With a 7B reader, L2R outperforms all LALMs we compare against on SAKURA and trails them by 6-12 points on MMAU and MMAR, despite training about 1,400x fewer parameters on orders of magnitude less audio data. However, because any LLM can serve as the reader, we show that a stronger reader narrows this gap without retraining any audio component. A new domain is added with one small head: with five labelled clips per species, it outperforms QLoRA fine-tuning of an LALM on the same clips by 13-26 points.

Bioacoustics and ethology encompass a wide range of audio understanding tasks, many of which stand to benefit from recent advances in large audio-language models. However, progress in the field has so far been assessed on a narrow set of tasks, primarily centered on label-centric biological category recognition, such as species and call-type classification. In this work, we introduce BEANS-Next, a benchmark grounded in a taxonomy of bioacoustics tasks spanning acoustic perception, biological category recognition, scene understanding, and in-context learning. Using BEANS-Next, we show that existing models exhibit limited performance beyond the task families emphasized by existing evaluations, constraining their usefulness for broader bioacoustic applications. To support progress on this broader task space, we also introduce ROOTS, a large-scale training resource built from expanded curated real-world data and previously underused behavioral and acoustic metadata, supplemented by audio-derived information and scalable synthetic generation where labeling is insufficient. We demonstrate that training on this dataset yields substantial progress across all task groups of BEANS-Next, moving audio-language models closer to their potential as general-purpose assistants for bioacoustics and ethology. To accelerate progress in the field, we open-source our benchmark, dataset, and data pipelines.

Language-queried audio source separation (LASS) enables extracting any sound source using natural language. However, adapting LASS models to application-specific sound classes is challenging due to noisy training data and limited semantic coverage of the CLAP-based control signals. We propose a framework comprised of an automated data engine for training-data curation and a two-stage optimization process for class-specific CLAP control signals. Our objective evaluations across seven sound classes show that data refinement and control signal optimization consistently improve source separation performance. Subjective evaluation with 17 participants further demonstrates perceptual improvements of the model trained with optimized control signals over baseline and similar commercial models.

Audio language models must handle dozens of distinct skills, from pitch comparison and speaker counting to musical tempo estimation and emotion recognition. Joint training on all skills at once causes interference: gains on one skill often come at the cost of another. We propose \textbf{SkillFormer}, which decomposes audio understanding into skill-specific low-rank adapters and composes them at inference time through a learned router. The router examines the question to decide which adapters to activate and how much weight each should carry, so that a pitch query engages different parameters than a genre classification query. An alternating training schedule updates each adapter on its own skill cluster before jointly calibrating the router, preventing the gradient conflicts that arise in standard multi-task optimization. SkillFormer adds fewer than 4% of the base model's parameters and requires no changes to the audio encoder or language backbone. Evaluated on three architecturally distinct models across MMSU, MMAU-Pro, and MMAR, it raises the average accuracy by 2.5 to 4.1 points, with balanced gains across perception, reasoning, and semantic subcategories.

Embodied, ego-centric intelligence fundamentally requires the ability to comprehend spatial audio within complex environments. While large audio-language models excel at mono-channel reasoning, they lack spatial awareness, discarding critical spatial cues that enable sound localization and that can improve the disentanglement of overlapping sound sources. To address this, we present SEA-LM, a Spatial Audio Understanding model. First, we introduce FOACODER, a layout-flexible spatial audio encoder trained on source localization and ego-centric voice activity detection objectives to encode First Order Ambisonics derived from variable-count, variable-position smart-glasses arrays via beamforming. We train a Multimodal Large Language Model (MLLM) to understand these spatial audio embeddings through a two-stage curriculum spanning six tasks, including sound localization and spatially selective transcription in settings with multiple speakers and overlapping sounds. To prevent the transcription outputs from dominating the next token prediction loss and overwhelming the direction predictions, we introduce a Spatio-temporal Weighted Cross-Entropy Loss. On our evaluation set, SEA-LM achieves lower azimuth and elevation MAE, higher temporal IoU, lower external-source hallucination and missing-source rates, and lower WER on most transcription tasks than compared baselines, while remaining robust across 1,211 smart-glasses array configurations with 4 to 9 microphones.

Audio decisions often depend on evidence that a transcript does not preserve, while their probability estimates can depend on how answer options are ordered. AudioJev maps a waveform, question and supplied alternatives directly to a candidate distribution through one shared full-parameter model. We define order calibration as preserving an answer's probability under meaning-preserving option permutations. Random-derangement SKL training pairs each question with a reordered view in which every alternative changes position, supervises both answers, and aligns the two distributions before applying a symmetric KL penalty. Inference retains a single candidate-scoring forward, with no calibration head or order ensemble. Across three training seeds, AudioJev reaches 68.88%/55.33% mean accuracy on complete MMAU/MMAR and reduces random-order SKL by 43.8%/60.5% relative to the single-view removal ablation. The same model handles intent, environmental sound, note properties, speech activity and conversational transitions. Paired removal ablations and multi-order evaluation measure predictive accuracy and probability stability together, establishing a direct audio interface whose calibration objective acts on candidate meaning rather than presentation position. Inference code and model weights are available at https://github.com/SihanLv/AudioJev-Inference and https://huggingface.co/shlv/AudioJev.

Large audio-language models (LALMs) are sensitive to input perturbations, such as noise, waveform corruption, and adversarial injections. We propose AnchorPrompt, an efficient adaptation method that keeps the model frozen and learns a single block of prompt vectors inserted at the decoder input, between the audio and question embeddings. We train these vectors through self-distillation over diverse audio and text perturbations. To improve answer consistency and mitigate hallucination, we use the model's prediction on the clean recording as the target for answerable inputs, and assign a refusal target when the audio lacks sufficient evidence to answer. Furthermore, AnchorPrompt is perturbation-agnostic at inference, requiring no prior detection of perturbations and enabling zero-shot transfer to unseen distortions. We evaluate three LALMs across three benchmarks and show that AnchorPrompt improves answer consistency in most tested conditions. Clean accuracy improves in six of nine model-benchmark pairs, with minimal impact on the remainder of 1.2% at most. Crucially, AnchorPrompt reduces hallucinations under severe audio corruption while keeping false refusals on clean audio rare. Finally, these consistency gains transfer to unseen perturbations, such as choice permutations and reverberation.

Audio language models (ALMs) are increasingly used for audio deepfake detection (ADD), yet existing benchmarks assess their verdicts or rationale plausibility without verifying the underlying acoustic evidence. To address this issue, we first introduce spoofing evidence-grounded audio reasoning (SEAR), a four-task AQA benchmark to evaluate ALM-based ADD through acoustic evidence identification and quantification, deepfake detection, and forensic rationale generation. We further propose a bona-fide-based acoustic evidence agent (BAEA), which equips a frozen ALM with controlled acoustic tools under \textsc{fixed} or \textsc{adaptive} evidence-acquisition policies. Experiments with six ALMs reveal a clear gap between plausible rationales and verifiable acoustic evidence reasoning, while BAEA-\textsc{Fixed} improves final verdicts and forensic rationales on both evaluation partitions. Controlled interventions further show that misleading evidence degrades both detection and grounding performance.

Audio deepfake detection (ADD) must remain effective when new spoofing attacks emerge after deployment. Emerging audio language model (ALM)-based ADD methods are built on predefined supervision from ground-truth labels or verified forensic rationales. However, this paradigm overlooks an ALM's own mistakes, which indicate where targeted supervision is most needed. To this end, we first introduce evolving spoofing environments for ALM-based ADD, where a new attack becomes dominant while previously observed attacks persist. Motivated by the above learning-from-mistakes perspective, we further propose SE-ADD, a self-evolving framework that iteratively adapts an ALM via low-rank adaptation (LoRA) using mistake-driven supervision built from its verdicts and self-generated forensic cues. All training samples receive direct authenticity supervision, while misclassified ones receive additional cue-augmented supervision. As verdicts and cues are regenerated by the updated ALM, the resulting supervision evolves accordingly. Experiments on two ALMs demonstrate the effectiveness of SE-ADD in generalizing to unseen attacks, reducing the equal error rate (EER) from $36.72\%$ to $7.52\%$ for Qwen2-Audio and from $19.93\%$ to $3.97\%$ for MOSS-Audio.

A large audio language model (LALM) turns a minute of speech into 750-1,500 tokens and prefills every one. Image-token pruning often cuts after the language model's first layers, where image tokens draw little attention. Audio tokens draw much more attention there, and their ranking is still far from final, so audio needs a ranking before the language model runs. Surprisingly, the attention an audio token will receive across the language model is already linearly predictable from its encoder output, before the language model runs. A linear map, fitted in closed form without labels, predicts this all-layer attention ranking at $ρ\geq .69$ on eleven of thirteen LALMs. Our method, Triage, cuts audio tokens by this prediction and, on multiple choice, cuts again at layer 2, correcting the prediction with the attention observed there. Triage sets its compression without labels, under two budgets that limit how far its output may differ from the model's own full-audio output. At the conservative budget, its word error rate and accuracy stay within .04 of full audio. At the aggressive budget, Triage beats every baseline in all twelve transcription cases. On multiple choice, at 2.2-5x compression, it outperforms DART, the strongest baseline on average, by .043 in mean accuracy. Because it cuts before the language model, it raises the audio that fits in Qwen2.5-Omni-3B's context window from 21.8 to about 62 minutes. At its most compressive point, Triage lets one GPU serve 4x as many concurrent 5-minute streams of that model. Project page: https://audio-triage.github.io

Speculative decoding accelerates autoregressive generation by using a smaller drafter to propose tokens for batched verification by a larger target. However, conventional speculative decoding couples drafting to the target's evolving verified prefix, serializing drafting and verification. We ask whether this dependency is necessary for source-conditioned generation. Our key observation is that, for audio language models, the input audio and user request can provide useful speculative candidates without following the target's evolving text prefix. We propose AS$^2$D (Audio Speculative Speculative Decoding), which enables target-decoupled drafting: an audio-conditioned drafter follows its own generation history while the target independently verifies and corrects ready candidates. Without usable candidates, the target advances alone. Thus, target feedback determines which candidates are committed but no longer determines when the drafter can make progress, enabling drafting and verification to proceed concurrently while retaining target-side verification and correction. We implement AS$^2$D in MNN for Android and evaluate two target models across four phones, seven datasets, and three tasks covering 12.2 hours of audio. Across four phones, AS$^2$D improves pooled ASR throughput by 42-76% over target-only decoding, while only 5.7% of evaluation windows are slower than target-only, compared with 58.1-63.0% for speculative baselines. For ASR, AS$^2$D reaches 97.33-98.20% of a hindsight per-window oracle's pooled throughput over the evaluated drafter/budget catalog. Native on-demand execution with a 7B target achieves up to 78% higher throughput than target-only. These results show that source-conditioned audio generation can relax the conventional dependence of speculative drafting on the target's evolving output prefix, exposing substantial parallelism for efficient inference.

Large audio-language models (LALMs) perform strongly on individual audio tasks, but whether these capabilities can be reliably composed remains underexplored. We conduct a controlled diagnostic study of capability composition in LALMs, requiring models to integrate audio-attribute recognition, cue-conditioned segment selection, and downstream ASR or question answering. We construct two-utterance inputs with distinct acoustic cues to evaluate composition across environmental sound, gender, and emotion cues, with ASR, Math QA, and Factual QA as downstream tasks. Across four open-source LALMs, compositional QA accuracy decreases in 39 of 40 model-task-cue settings, by an average of 26.7 percentage points. ASR exhibits a similarly consistent degradation, with WER increasing in 39 of 40 settings by an average of 28.5 percentage points, while the magnitude of degradation varies across models, cue types, and cue salience. We further probe these failures through output format, positional preference, and chain-of-thought (CoT) analyses. Our study reveals a systematic gap between possessing individual audio capabilities and reliably composing them.

Audio large language models (ALLMs) can reason about the content of audio recordings to perform complex tasks. However, these capabilities usually collapse in real-world environments when background noise and competing sources mix the target sound. Inspired by long-term memory in human listening, we propose Long-Term Memory-Guided Audio Enhancement (LTM-AE) to improve selective target perception by refining the audio representations of ALLMs without training. LTM-AE extracts representations in hidden states from separate clean reference recordings as long-term memory for each category, guiding enhancement toward a user-specified listening target. We reconstruct incoming audio tokens in the selected category long-term memory and interpolate the reconstructions with the original tokens before language backbone decoding. This interpolation controls the influence of stored auditory experience while keeping all ALLM parameters fixed. Diagnostic readouts across twenty sound categories and three ALLMs show that LTM-AE strengthens responses to a specified target amid three interfering sources. Averaged over constrained and free-form classification, accuracy gains over raw mixtures range from 29.53 to 46.15 percentage points across multiple open source models. For speech content recovery, LTM-AE with an additional learned token-level gate reduces Qwen2-Audio's word error rate from 23.07% to 14.77%. This work takes an initial step toward using principles of human long-term memory to enhance ALLMs for real-world listening. Our code is available at https://github.com/aynlp/ltm-audio-code

We introduce HEAR (Human-recorded Evaluation of Audio-LLM bias by Real speakers), a large-scale, ecologically valid benchmark comprising 87k real human audio samples from 843 demographically diverse participants. HEAR enables comprehensive evaluation through Multiple Choice Question Answering (MCQA) and open-ended long-form tasks. To our knowledge, this is the first large-scale voice benchmark grounded entirely in authentic human speech. We evaluate model behavior across both real-time speech-to-speech and speech-to-text architectures. Our results reveal that voice-conditioned bias is a model-specific property. Furthermore, we demonstrate that personalization instructions consistently exacerbate demographic disparities. Our findings establish that voice bias is a controllable model characteristic, providing a foundational framework for future bias mitigation and evaluation in Audio-LLM development.

Large audio-language models (LALMs) excel at understanding and reasoning tasks over atomic sound events, yet their ability to infer higher-level human activities from such fine-grained events remains largely unexamined. Everyday human actions and activities, such as setting a table, cleaning the house, or preparing a breakfast emerge compositionally from temporally distributed sound events, requiring abstraction beyond the event-centric granularity that dominates current training and evaluation paradigms. Our benchmark evaluates a wide set of LALMs under a principled framework that tests how language-based reasoning, grounded in acoustic perception, structures sound abstractions into higher-level understanding. By systematically varying exemplar typicality and distractor similarity, our evaluation exposes \added{that current models do not reliably perform compositional inference from atomic acoustic events to higher-level human activities solely from audio.} All data, taxonomies, and evaluation scripts are publicly available on our companion website: https://alm-sounding-actions.onrender.com/

Reasoning about temporal structure of audio recordings requires Large Audio Language Models (LALMs) to associate sound events with their temporal position. Understanding the underlying mechanisms is a first step toward diagnosing failures and identifying model components that may need improvement. Using mechanistic interpretability, we investigate how temporal information is represented and bound to sound events in three open-source LALMs. We find that across all three, event-specific location becomes concentrated in event name representations at intermediate modality integration layers. These representations encode coarse event position along a low-dimensional, curved relative time trajectory. Steering event name representations along this trajectory systematically shifts before/after beliefs, providing evidence that these representations contribute to coarse temporal reasoning. In contrast, the same interventions do not reliably shift predicted onset timestamps, suggesting that coarse temporal reasoning and precise metric event localization rely on distinct mechanisms.

Spatial audio large language models (LLMs) enable embodied agents, wearable assistants, and immersive systems to recognize sound events, localize sources, and reason about their spatial relationships. However, existing spatial audio LLMs often rely on early fusion of acoustic and spatial features and source-agnostic token representations. These designs make it difficult to preserve the correspondence between individual sound events and their spatial attributes, particularly in multi-source scenes. To address this limitation, we propose SAIL, a Spatial Audio Intelligence framework with LLMs that preserves acoustic-spatial structure and source-level correspondence from audio encoding to LLM alignment. SAIL introduces a Disentangled Spatial Audio Transformer that represents Mel-spectrogram and interaural phase difference features as separate acoustic and spatial streams. Source-discriminative task queries further learn event, direction, and distance information for each source. A Dual-Stream Q-Former then aligns the two streams with the LLM using acoustic and spatial queries organized by source slots. Compared with the early-fusion baseline, SAIL achieves consistent improvements in dual-source sound event detection, direction and distance estimation, and spatial reasoning. These results demonstrate the importance of structured, source-discriminative audio representations for multi-source spatial understanding and reasoning.

Large audio-language models (LALMs) produce fluent responses about audio but often hallucinate by making plausible yet ungrounded claims. Existing audio hallucination benchmarks mainly measure response correctness, leaving it unclear whether an LALM hallucinates or simply fails to understand the audio. We challenge correctness-based evaluation by defining two hallucination categories: (i) context, where claims are not grounded in the audio; and (ii) knowledge, where claims about audio-related topics lack support from externally verifiable facts. We introduce MISHAP-Bench, a comprehensive benchmark with 12,000 challenging open-ended question-audio pairs and a rigorous evaluation pipeline covering both categories. To evaluate open-ended responses, we propose a groundedness judge that uses reference rubrics and judge prompts guided by human annotations. We extensively evaluate ten state-of-the-art LALMs and show that hallucination remains substantial. Even a frontier model such as Gemini 3.7 Flash reaches a hallucination rate of 36.5%. We further adapt and benchmark four mitigation methods from multiple domains for LALMs. Despite some improvements, effective hallucination mitigation remains an open challenge. Finally, we call on the community to evaluate hallucination and benchmark mitigation methods with MISHAP-Bench.

Multimodal large language models answer audio questions, but how they represent auditory semantics and use them in decisions remains unclear, limiting our understanding of response formation. We study dog barking in Qwen2.5-Omni-7B using Jacobian lens (J-lens) readout and directional interventions. We define the dog direction as a J-lens-derived hidden-state vector associated with dog; adding or removing its component modulates dog-related information. We find this information decodable without dog/bark prompt cues or animal-identification requirements. Directional interventions change response tendencies and some final answers, with effects concentrated in late-layer states immediately before generation across species classification, vocalization classification, and sound description. The dog direction shows no comparable advantage over controls in animal/other classification. These results provide causal-intervention evidence that the dog direction affects output scores in a task-dependent manner, most consistently at L22 and L24 immediately before generation.

Audio large language models allow users to interact with the model through speech. When an input recording is too degraded, the model may misinterpret the user's query and respond based on an incorrect transcription. In this paper, we study model-conditional transcription reliability: whether an Audio LLM can recognize when its own transcription is unreliable. We first prompt the Audio LLM to assess whether its own transcription would be reliable, and find that the model is a poor judge of its own transcription reliability: in most cases, it predicts that its transcription will be reliable. We find that existing approaches, including speech quality predictors, audio LLM generation uncertainty, and transcript-conditioned WER estimation, provide limited signals for detecting transcription failures. In contrast, we discover that transcription reliability is strongly represented in the model's audio-encoder representations. Based on this observation, we devise a lightweight reliability predictor that operates on representations extracted by the frozen audio encoder and predicts the reliability class before generation. The reliability predictor can trigger a clarification request from the user when their voice query is predicted to be unreliable, while allowing reliable queries to proceed without modifying the underlying Audio LLM. Our predictor achieves 81.10% in-domain and 78.09% cross-domain macro-F1 scores, outperforming the strongest baselines by 10.33 and 11.93 points, respectively. Finally, we show that reliability labels can transfer across Audio LLM families, and that transfer performance is closely related to the alignment of their model-specific reliability boundaries.

Audio language models pass speech and text through a single decoder. We ask whether that decoder represents a distinctive feature in the same direction when a phoneme is heard and when it is read. For minimal pairs of phonemes differing in one feature, we take the offset between the two members' mean representations. Averaging those offsets gives a direction for each stream, and we measure the cosine between the two. Because the two streams already agree about arbitrary phoneme pairs, we compare every measure against a reference built from random pairings rather than against zero. We apply this to 6 models, 7 features and 15 languages from 11 families. Only voicing in the two Qwen2.5-Omni models exceeds that reference after correction for multiple testing, and the reference varies by a factor of seven between models. In three of the six models, voicing has one direction in audio across the 14 languages with enough minimal pairs to measure it, and every language pair agrees in two of them. The model family, not the model size, predicts which stream represents a feature.

Natural conversations make both speech recognition and speaker attribution challenging for ASR, as speakers take turns, overlap, and reappear over time. We propose STAM-ASR, Speaker-Temporal Anchoring with Memory, a lightweight framework that extends an already pretrained AudioLLM for multi-speaker ASR. Without relying on an external diarization system, STAM-ASR learns speaker activity and speaker-aware representations directly from intermediate AudioLLM features. Hence providing explicit who and when cues to modulate the AudioLLM's semantic representation without explicit speech separation. STAM-ASR further maintains fixed-size speaker and conversational memories to carry complementary context across turns. We evaluate STAM-ASR on AMI, ICSI, LibriCSS, and NOTSOFAR-1 across close-talk, far-field, overlapping, and cross-domain conditions. Our reported results shows that speaker-temporal conditioning and memory provide complementary benefits, while the gap between reference and predicted speaker activity identifies robust speaker tracking as a key remaining challenge.

Large audio-language models (LALMs) expand language models to process and interpret audio, but also expose them to heterogeneous audio jailbreaks. We ask whether successful jailbreaks reflect failures to recognize harmful intent or failures occurring after such recognition. Layer-wise probing reveals the latter: risk-related information remains decodable from intermediate representations, yet the internal risk signal fails to translate into refusal in later-layer processing. We identify this discrepancy as the risk-to-refusal gap. Building on this finding, we propose AEGIS, a detect-then-intervene defense whose mid-layer risk gate selectively activates downstream safety adapters. Across six LALMs and three heterogeneous audio jailbreak benchmarks, AEGIS reduces the average unsafe rate from 17.9% to 0.4%, while causing only a marginal increase in over-refusal on benign inputs. These results establish selective internal intervention as an effective path toward more robust refusal in LALMs. The code is available at https://github.com/azzzzliao/aegis-audio-defense.

Audio-language models can produce confident answers unsupported by the audio, motivating uncertainty estimates that identify unreliable responses. We compare probability-based, sampling-based, self-verification, evidential, and contrastive measures across four open-weight models and five audio QA benchmarks. In multiple-choice evaluation, first-token measures are strongest overall, with top-1 probability achieving a mean AUROC of .740, compared with .708 for ten-sample discrete semantic entropy, while requiring no additional model calls. Across four benchmarks, shifting from multiple-choice to open-ended evaluation lowers mean accuracy from 57.6% to 36.6%, yet uncertainty remains predictive of errors: semantic entropy, maximum token entropy, and semantic agreement achieve mean AUROCs of .697, .694, and .693, respectively. To test whether uncertainty reflects the evidence available to answer the question, we perform input ablations that remove either the audio or the question. Across top-1 confidence, entropy, and sampling-based measures, removing audio reduces error-detection AUROC by .101 on average, compared with .010 when removing the question. Together, these results establish efficient uncertainty baselines and show that uncertainty in audio-language models depends substantially more on available audio evidence than on question text.

Audio-language models (ALMs) can exploit textual shortcuts to answer questions while overlooking acoustic evidence, weakening audio understanding. On-policy distillation (OPD) trains compact ALMs by supervising student-generated responses with teacher predictions, but does not explicitly distinguish acoustic support from linguistic predictability. We propose Reward-Tilted On-Policy Distillation (RT-OPD) to strengthen acoustic grounding. Given the same question and student-generated text, a frozen teacher predicts the next token with and without audio inputs. Their log-probability contrast defines a reward that reshapes the teacher distribution for reverse-KL distillation, emphasizing the additional evidence provided by audio. Across two compact students and three benchmarks, RT-OPD consistently outperforms Vanilla OPD. Experiments with silenced and replacement audio further suggest that RT-OPD strengthens the student's reliance on acoustic evidence. Our 3B model achieves 72.72% accuracy on MMAU, the highest among the compared 3B models and competitive with several 7B and 8B models. Code and model checkpoints are available at https://github.com/KaiyangLi1992/RT-OPD.

Self-supervised learning (SSL) countermeasures (CMs) have shown strong performance in recent years. However, they often show degraded performance while facing unseen spoofing attacks and mismatched conditions. This study examines the Voxtral audio-language model (ALM) framework for spoofing detection, as a step toward combining CM capabilities within the ALM framework. We analyze how Voxtral captures spoofing cues through audio-text processing and propose an instruction-guided approach that uses label-sequence likelihoods to evaluate bonafide and spoofed speech. Experiments on the ASVspoof databases show that without task-specific adaptation, the LLM layers emphasize semantic representations, reducing the separability of spoof-discriminative acoustic cues compared to the Whisper-based audio encoder. Consequently, spoofing-related information becomes less separable after language-model processing. We also applied lightweight adaptation using weight-decomposed low-rank adaptation (DoRA) to the Voxtral model and propose the Spooftral model, achieving an equal error rate (EER) of 4.25% on the ASVspoof5 evaluation set.

Audio-language models (ALMs) integrate acoustic perception with the knowledge encoded in language models, enabling contextual understanding of auditory events. Making these capabilities practical on devices with limited memory and computation motivates our focus on small ALMs with fewer than 200M parameters. We introduce a recipe that brings together architecture, data, and three-stage training to build Mizar, a 159.3M-parameter ALM. Its architecture connects a compact CED-Small audio encoder to SmolLM2-135M through a frequency-merging mapper. With supervision drawn from ReasonAQA, AudioMCQ, and AVQA, the model undergoes three training stages: audio-language alignment (Stage 1), audio-dependent fine-tuning (Stage 2), and post-training (Stage 3) aimed at strengthening weak skills while retaining learned capabilities. Across five random seeds, Mizar achieves mean accuracies of 52.92% on MMAU, 42.42% on MMAR, and 36.02% on ADQA-clean, surpassing the previous best-performing ALM below 200M parameters on all three benchmarks. It also supports local inference on a single CPU: on questions from the MMAU benchmark, the mean latency from opening the audio file to generating a complete answer is 1.09 seconds. Code and checkpoints are available at https://github.com/KaiyangLi1992/Mizar_159M.