Audio-Language Model Evaluation

Voice products increasingly need affective cues that are present in speech but absent from transcripts. We introduce VocalAffectBench, a public, test-only benchmark for evaluating whether AI audio models can identify expressed vocal emotion from raw audio. The benchmark contains 273 human-recorded English WAV clips from 51 speaker accounts totaling 1.95 hours across seven labels: angry, disgusted, fearful, happy, neutral, sad, and surprised, with 39 clips per class. All baselines are evaluated from audio alone, without transcripts or contextual metadata. Across six released baselines, average accuracy is 35.5%. The strongest baseline, gemini_3_5_flash, reaches 46.5% on the seven-way task, above the 14.3% random baseline but far from robust emotion recognition. A secondary valence-bucket analysis maps labels into positive, neutral, and negative classes, excluding surprised because its valence is ambiguous. Aggregate accuracy under this coarser view is 50.9%. Performance is highly uneven across classes. By recall, neutral is identified most reliably at 75.6% averaged across baselines, while surprised and fearful reach only 10.7% and 15.4%, respectively. These results show that the evaluated baselines can extract some affective signal from speech, but discrete expressed-emotion recognition remains fragile, especially for non-neutral emotions that are often most important in voice agent workflows.

Speech and audio LLMs are evaluated by comparing waveform predictions with predictions from an automatic speech recognition (ASR) transcript. For fixed closed-set tasks, this conflates acoustic evidence with the need to invoke a generative audio model. We estimate incremental call value with matched selectors sharing pre-call evidence. Each policy may retain the transcript label, use a local encoder, or invoke a generative model; matched control removes generative actions but preserves pre-call evidence and development selection. On VocalSound, transcript-only accuracy is 0.296, while supervised CLAP and WavLM controls reach 0.850 and 0.854 without calls. Full selector reaches 0.925 at 12.5% calls versus 0.921 for matched No-call selector (difference 0.004; 95% CI [-0.025, 0.033]). Thus, results do not show a call gain after transcript and encoder evidence are available. Relevant quantity is incremental accuracy from allowing calls, not the waveform-transcript gap.

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000--0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919--0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

General audio comprehension now covers speech, sound, and music over durations from seconds to hours, driven by large audio-language models (LALMs) that are increasingly omni-modal. Yet the benchmarks that test them still rely on clips of seconds, where scores saturate and models converge; recent long-form efforts extend duration but evaluate long audio much as short clips are. We introduce LongAudioSpan, a benchmark that spans both duration and depth: it pairs audio from 10 minutes to over 2 hours with 3,240 questions across three cognitive levels, namely perception, understanding, and reasoning. Two paths supply the questions, differing in how question content is sourced and how ground truth is obtained. Native QA extracts questions from the audio's content, posing each as a multiple-choice item and an open-ended one graded by detailed rubrics. Anchor QA instead injects ground truth, planting acoustic anchors into the audio and building a perception-to-reasoning chain scored only to the first error. A fully automated pipeline constructs every item through structured captioning, QA generation, and adversarial critic feedback. Evaluating 12 LALMs on LongAudioSpan, we find the hard part comes before reasoning: distilling a few relevant facts from a long, redundant signal. This difficulty grows with audio length and falls hardest on perception, especially temporal grounding. LongAudioSpan is available at https://huggingface.co/datasets/holvan/LongAudioSpan.

While text-based hallucination detection is well studied, reference-free detection of factual alterations in speech remains underexplored, especially for low-resource languages. Our spoken benchmark comprises 12,013 English, Russian, and Kazakh news samples with three synthetic alteration types and three severity levels, pairing source articles with rewrites as text, synthesized audio, and ASR transcripts. We add 290 fact-checked misinformation items collected in Russian (225) and Kazakh (65), translated into the other language and rendered through the same TTS-ASR pipeline. We evaluate fine-tuned multilingual encoders and zero-shot multimodal decoders on text, transcripts, and audio. Detectors receive only target inputs without source articles or external evidence; the task evaluates reference-free classification rather than evidence-grounded verification. Encoder degradation from source text to transcripts generally tracks per-language ASR error on the binary task. Among decoders, only Gemma-3n exceeds the binary majority-class baseline in macro-F1, on transcripts only; the other four fall below their respective baselines. Comparisons between audio and transcripts are confounded by differences in evaluation coverage and class balance. Synthetic-trained detectors achieve 0.82--0.88 macro-F1 on real-world misinformation source text; Russian provenance analysis reveals veracity-related and model-dependent machine-style signals, a key confound in synthetic hallucination benchmarks.

Large audio-language models (LALMs) have shown promising progress in understanding speech, music, and general sound events, yet their ability to reason about how audio signals are degraded remains underexplored. Existing benchmarks primarily evaluate semantic understanding, event recognition, or high-level audio reasoning, leaving a basic question unanswered: Do LALMs understand the differences in audio quality? We introduce MRMAD, a Multi-Round Multi-Audio Degradation benchmark for evaluating audio degradation perception and understanding in LALMs. MRMAD spans speech, music, and sound, and frames evaluation as multi-turn dialogues across multiple audio inputs, requiring models to identify types of degradation, compare severity, and perceive corruption changes across turns. Unlike current single-turn audio-language benchmarks, MRMAD evaluates whether LALMs can maintain consistent degradation hypotheses with new evidence and comprehend low-level acoustic phenomena over multi-turn dialogues. Through a systematic evaluation of 18 representative LALMs from non-thinking to reasoning and Omni models, we find that current models often recognize coarse content while failing to diagnose, compare, or reason about degradations reliably. Human evaluations further reveal a significant perception gap between LALMs and human listeners. MRMAD thus exposes a critical yet overlooked aspect of audio-language understanding and provides a diagnostic foundation for building future LALMs that are robust to real-world acoustic conditions.

Recent advances in large language models (LLMs) and multimodal LLMs (MLLMs) have created new opportunities for wearable speech interfaces, with smart glasses providing an egocentric platform for continuous audio sensing and assistance. However, speech recognition and understanding in this setting remain challenging because of dynamic acoustic conditions, speaker overlap, and the spatial ambiguity introduced by wearer-centered recording geometry. To support systematic evaluation in this setting, we introduce the IEEE SLT 2026 SmartGlasses Challenge for egocentric multi-speaker speech processing. The challenge consists of two tracks, Dyadic Dialogue Understanding and Multi-party Meeting Understanding, and jointly evaluates Time-Stamped Speaker-Attributed Automatic Speech Recognition (TSA-ASR) and Spoken Language Understanding (SLU). It is built on a 106-hour four-channel egocentric speech dataset containing 714 sessions collected in real-world scenarios. This paper describes challenge tasks, dataset construction, submissions, and summarizes the main findings from the shared evaluation. The results show that heavy speaker overlap remains a major factor affecting TSA-ASR performance, while paralinguistic acoustic understanding continues to be difficult for current audio-language models in complex SLU settings. Further details can be found on the official challenge website.

Automated Text-to-Speech (TTS) evaluation methods (Mean Opinion Score (MOS) predictors and Audio Large Language Models (Audio-LLM) judges) are expected to reflect human perception, yet it is unclear how well they capture the distinct aspects of speech that listeners actually perceive. We deconstruct "naturalness" into a linguistically grounded annotation schema spanning 10 distinct perceptual dimensions, and use it to construct the first dimension-level meta-evaluation benchmark for TTS, comprising 860 utterances annotated by trained linguist raters. Results from benchmarking four MOS predictors and four Audio-LLM judges reveal that MOS predictors collapse onto acoustic signal quality, while Audio-LLM judges show selective, prompt-dependent detection that does not generalise across all dimensions. Neither class reliably captures a breadth of linguistically structured speech errors. Our dataset, annotation schema, and evaluation code are publicly released to support more targeted and interpretable TTS evaluation.

Large audio-language models (LALMs) have demonstrated strong capabilities in understanding diverse audio inputs. This diversity includes low-frequency signals that are inaudible to humans but can still enter the model and influence its generation. However, the practical impact of such low-frequency inputs on LALMs remains largely unexplored. In this paper, we propose Intermittent Low-Frequency Lockout (ILL), an inaudible red teaming method that evaluates this risk using a universal waveform template in a black box setting. ILL uses Sentence Attention Scale Estimation to determine active intervals and Frequency Confusion Transfer to construct a low-frequency waveform with continuous phase from corpus spectral variation. To mitigate this risk, we propose Distributional Requery Guard (DRG) to detect low-frequency distribution shifts and conditionally request a second recording for semantic recovery. Across six LALMs and multiple audio understanding tasks, ILL reduces accuracy by up to 67 percentage points while receiving a mean human audibility rating of 1.33, close to 1.17 for clean audio; DRG raises mean attacked accuracy from 28.5% to 46.1% after clean reacquisition. These findings identify a previously overlooked safety risk for LALMs and provide a foundation for future research on robust audio understanding.

Audio-language models (ALMs) are increasingly used as judges for speech-to-speech systems, but a judge that receives audio may not actually use paralinguistic evidence. We introduce counterfactual audits for paralinguistic response evaluation. Each audit item holds the transcript fixed while varying affect, prosody, or the timing of an affective shift, forcing a valid judge to track the audio cue rather than lexical content or response style. We evaluate ALM judges using a native one-context judgment protocol and a contrastive recoverability control, then further decompose each item into its constituent perception and response-mapping skills. This yields useful diagnostic states that identify different sources of judge failures. Across Gemini, GPT, and open audio models, we find that contrastive success often overstates native judge reliability, and that similar aggregate accuracies can hide different failure modes. These results suggest that ALM judges should not be evaluated by accuracy alone, instead requiring thorough behavioral audits before deployment.

Spoken Language Understanding (SLU) is the core component of task-oriented dialogue systems and a pivotal link in achieving seamless human-agent interaction. While traditional SLU can effectively extract user semantics for closed-set tasks after in-domain supervised fine-tuning, it faces significant challenges in leveraging in-context learning for open-domain tasks due to its ambiguous rule definitions. This work proposes Spoken Function Calling (SFC), a novel semantic understanding perspective that optimizes semantic understanding with structured rule definitions, to evolve beyond traditional closed-set SLU. Specifically, we curate and extend a suite of spoken functions based on traditional SLU datasets, construct a multi-agent system to synthesize the SFC-Bench dataset, evaluate the performance of Large Language Models (LLMs) and Large Audio Language Models (LALMs), and enhance the SFC capabilities of LALMs through post-training. Experiments demonstrate that SFC outperforms traditional SLU, substantially enhancing the semantic extraction accuracy for LLMs and LALMs.

Speech language models are increasingly evaluated on paralinguistic tasks by the accuracy of prompted answers, but answer accuracy combines failures at different stages of the audio-to-answer computation. We introduce a generation-aligned diagnostic ladder that compares the emitted answer, the option logits, an affine readout of those logits, and a linear readout of the hidden state at the same answer token. Successive differences separate endpoint, decision-rule, and readout-coverage gaps. Across five systems and two emotion corpora, state decoding exceeds generation by 27.8 accuracy points on average, and both the decision-rule and readout-coverage gaps are positive in all ten conditions. A label-free logit correction improves generated accuracy in every condition, showing that part of the decision-rule gap is actionable. In rank-matched comparisons, emotion information outside the native readout generalizes to held-out speakers and survives controls for measured acoustic descriptors, but replacing the selected readout-external directions usually has little effect on emitted answers. These results distinguish information availability from behavioral use and localize performance losses across the decision rule and the state-to-answer readout.

We benchmark eleven audio classification methods: five task-aware closed-set LLMs (four Gemini models plus open-weight Kimi-Audio-7B-Instruct), four fixed-vocabulary taggers (YAMNet, PANNs, Whisper-AT, and SSLAM), a zero-shot audio-text model (CLAP), and an audio-grounded LLM (BAT). We evaluate them on a closed-set sound-source identification task over 2,242 clips spanning 23 fine-grained classes and 11 categories. Since these methods differ fundamentally in how they receive the task and how outputs are scored, we group them into four evaluation tiers rather than one leaderboard, reporting macro Precision, Recall, F1, and false-negative rate per tier. The best model, Gemini-3.1-Pro-Preview, reaches 85.6 percent category-level F1 and 56.7 percent fine-grained F1. Kimi-Audio is competitive for its size, reaching 67.5 percent category-level F1 and 32.9 percent fine-grained F1, but fails to answer 1.6 percent of samples. SSLAM and CLAP match or exceed the best closed-set model at the category level without seeing the candidate list, but fall behind at the fine-grained level. Analyzing the Gemini models' chain-of-thought across 8,968 responses, we find that response length does not predict accuracy, an apparent "holistic judgment beats detailed analysis" effect is better explained as a difficulty confound, and wrong answers are stated confidently 92 to 100 percent of the time. We report full per-class confusion matrices and metrics for all eleven methods, identify the structural error modes behind most of the accuracy loss between granularities, and give practical guidance for choosing among these method families.

Unified audio models capable of audio understanding, audio generation and, increasingly, audio editing are proliferating rapidly. Yet a basic question about them remains unanswered: do the two heads of a unified model agree about the same audio? Current practice evaluates each capability in isolation on specialized benchmarks, and never asks whether a model can make sense of its own generations. We present TORUS, the first self-coherence test for audio-native unified models. TORUS comprises 48 three-stage self-coherence tests carrying 432 six-option questions spanning speech, sound and music across five task families. We holistically evaluate five open unified models alongside a Cascaded Baseline that combines state-of-the-art specialized generation, editing and understanding models. The best unified model answers 50.5% of questions against the Cascaded Baseline's 63.2% and a 16.7% chance floor. Models struggle on audio editing. Among the evaluated audio models (specialized and unified), we observe limited self-coherence, and thus position self-coherence as an essential test for future audio systems.

With the development of audio large language models (AudioLLMs), audio captioning needs to move from brief descriptions toward open-ended and fine-grained free-form descriptions. Existing evaluations often focus on generation quality or task performance, making it difficult to diagnose information coverage and description reliability. We propose MMAC, a \textbf{M}assive \textbf{M}ulti-dimensional benchmark for \textbf{A}udio \textbf{C}aptioning. MMAC contains 5,638 audio clips from more than 20 data sources, covering 6 capability categories and 15 evaluation dimensions. Given a model-generated caption, MMAC checks whether it mentions relevant information in the target dimension and whether the mentioned content is consistent with the reference label. We evaluate representative open-source and proprietary AudioLLMs. Results show clear differences across evaluation dimensions, information coverage, and description reliability. We will release the MMAC benchmark and evaluation code.

Audio-capable foundation models enable end-to-end spoken interaction, but they also introduce safety risks beyond transcript content. It remains unclear how much jailbreak capability can arise from matched-text variation in speech delivery rather than from lexical rewriting or broader style transfer. We study this question by holding transcript content fixed and varying six speech-delivery presets whose acoustic attributes may co-vary. We present PJ-Break, a black-box evaluation protocol with presets targeting arousal, authority, and speaking rate, together with AdvAudio-Prosody, a 600-sample benchmark with acoustically verified attributes. On the exact post-QC Qwen2-Audio panel, the Q=1 Panic (38/95), Anger (35/95), and Fast (32/95) presets are all well above Neutral (4/95). The fixed six-query pool covers 44/95 Qwen2-Audio seeds and 15/95 GPT-4o seeds and exceeds a matched-budget StyleBreak reimplementation (27/95) on Qwen2-Audio. A same-voice pool excluding the confounded Commanding condition still reaches 40/95, and a retained-panel ablation shows emotional-delivery audio alone (44/95) is far more effective than emotional text alone (11/95). Exploratory surrogate diagnostics and pilot mitigation observations are secondary, non-core analyses. Overall, matched-text speech delivery should be treated as a first-class factor in Audio LLM safety evaluation

As large audio language models (LALMs) advance, robust evaluation frameworks have become essential. In this context, Spanish speech understanding under realistic acoustic conditions has received particularly little attention. We introduce ESCUCHA, the first Spanish speech understanding benchmark designed to evaluate LALMs across heterogeneous acoustic conditions and reasoning abilities. ESCUCHA comprises 1,000 human-curated questions paired with audio, totaling 162.9 hours sourced directly ``from the wild'' rather than drawn from existing datasets, with durations ranging from a few seconds to over 80 minutes. The benchmark emphasizes reasoning, spanning 9 perceptual and 10 reasoning categories, and it captures linguistic diversity through multiple Spanish accents and non-normative speech. ESCUCHA further includes multi-audio questions, spoken questions, and audio instructions, and it flags which questions support open-ended evaluation. Benchmarking several state-of-the-art multimodal and speech models reveals substantial performance gaps relative to trained humans.

Recent advances in text-to-speech and voice cloning make high-quality spoofing inexpensive and scalable, threatening voice authentication systems, especially automatic speaker verification (ASV). Existing defenses mainly address this threat through binary countermeasures (CMs) for deepfake detection or spoofing-aware speaker verification (SASV), where current systems are dominated by modular ASV-CM fusion and cascaded pipelines. Although large audio language models (LALMs) have shown promise on related audio tasks, including CM and ASV, their use for SASV remains unexplored, despite their capacity to produce natural-language rationales for auditing and robustness beyond discriminative predictions. This work systematically evaluates LALMs for SASV against conventional pipelines under zero-shot prompting, supervised adaptation, reasoning-oriented training, and reinforcement-learning-based optimization. Our results show that pretrained LALMs are near chance in the zero-shot setting, confirming that they are not natively suited to SASV, but that task-specific adaptation closes this gap. We further find that competitive SASV performance can be achieved through several distinct routes. These findings position LALMs as a promising and auditable foundation for unified SASV, while clarifying where conventional cascade systems still lead.

Large audio-language models (LALMs) are increasingly used as automatic judges for speech evaluation. However, high agreement with human ratings does not guarantee that their verdicts are grounded in the audio. A judge may instead rely on specialist labels or reference data supplied by the evaluation protocol itself, taking a shortcut in place of listening to the audio. In this paper, we audit such protocol-level ``shortcuts'' in LALM judges across three common deployment protocols: feature-blueprint judging, where the audio is replaced by a structured text description of acoustic features, reference-conditioned judging, and pairwise A/B comparison. Across six judges and four attributes, we find that several LALMs rely on protocol-level shortcuts. For example, in feature-blueprint judging, incorrect specialist labels reduce five judges' emotion accuracy to 0.10 or below, and in concatenated A/B comparisons, Qwen3-Omni-Thinking often picks the same slot regardless of order swaps. These results indicate that aggregate agreement can overstate the validity of LALM judges unless the model and the evaluation protocol are assessed jointly, and that each model-protocol pair should be evaluated with a matched shortcut probe.

Recent text-to-audio models generate high-quality audio, but often fail to follow instructions involving multiple sound events and temporal order. This gap arises because existing evaluation and training signals mainly emphasize global similarity or perceptual quality, with limited supervision on instruction-level correctness. We propose an instruction-level framework that uses audio-aware large language models (ALLMs) as fine-grained judges to verify target event presence and temporal relations in generated audio. After validating ALLM judgments on benchmarks and through human verification, we use their feedback to construct preference pairs for direct preference optimization. We further introduce S3Bench, a narrative benchmark for evaluating multi-event temporal instruction following. Experiments show that our method improves event completeness, temporal ordering, and joint instruction-following accuracy across existing benchmarks and S3Bench, while maintaining audio quality.

Audio-language embedding models such as CLAP are widely evaluated on matching present sound events, but rarely on negation. We show this affirmation-only evaluation hides a key limitation: these models fail to encode negated sound concepts, mapping affirmative and negated captions to nearly identical representations. To expose this blind spot, we introduce NegEval-Audio, a framework that converts existing datasets into two negation-aware tasks, Retrieval-Neg and Multiple-Choice Negation (MCQ-Neg), to probe whether models distinguish present from absent events. On AudioCaps and Clotho, performance degrades sharply under negation, with negation-type MCQ accuracy falling far below chance, and the failure persists even for a recent multimodal LLM-based embedding model. While a training-free steering method improves MCQ-Neg, it yields marginal gains for Retrieval-Neg. This indicates that affirmation bias is a fundamental flaw in the representation geometry, necessitating explicit negation-aware training objectives.

Sound symbolism, the human tendency to map speech sounds to perceptual qualities such as roundness or sharpness, arises primarily from the acoustics of speech rather than spelling. Whether Speech Language Models (SLMs) share this tendency remains open, as prior evaluations rely on text or images rather than real speech. We study it using genuine human speech recordings, comparing model judgments against human data across the auditory, crossmodal, and visual components of the effect. We find that SLMs' auditory judgments align poorly with human perception and miss the acoustic cues, such as spectral tilt, that drive human intuitions, and open-weight models cannot reliably link a heard sound to its corresponding shape. With a visual-only control ruling out shape perception, the weakness localizes to how speech is represented, suggesting that perceptual alignment depends not on stronger vision but on speech representations that capture the cues humans hear.

We report the empirical reliability of Gemini models as audio judges that score full-duplex agent conversations directly from the raw stereo waveform, tested across three models in the Gemini family: 2.5 Flash, 3.5 Flash, and 3.1 Pro. Our primary evidence base uses Gemini 2.5 Flash as the ground-truth model, validated against three calibrated human raters on 209 stereo sessions, scored on 8 production dimensions: 152 full-duplex conversations across 13 accent-and-condition strata, together with 57 adversarial defect-injected clips. The evidence for Gemini 2.5 Flash is consistent across three tests. (i) On 5 of 8 dimensions the LALM-human Spearman rho departs from the pairwise human-human rho by at most 0.07, and on 7 of 8 dimensions the two quantities 95 percent bootstrap confidence intervals overlap. (ii) The LALM agrees with the three-rater human mean within 1 point on 60 to 92 percent of sessions on 6 of 8 dimensions. (iii) On 45 of 48 (defect, dimension) cells the LALM is as sensitive as humans or better under Newcombe-Wilson 95 percent confidence intervals, though most of these are underpowered nulls rather than demonstrated parity. Rank-ordering ability transfers across the Gemini family: 3.5 Flash improves simple agreement to 8 of 8 dimensions, while 3.1 Pro rates several dimensions markedly lower than humans despite comparable rank correlation. A model swap should be re-validated on calibration specifically, not assumed from rank-correlation alone. We identify four areas where deployment requires care, and we estimate that human rating alone for our current evaluation cadence costs roughly two orders of magnitude more than the equivalent LALM workload. The data presented here provides a defensible empirical basis for deploying the LALM as a substitute or fourth rater on the dimensions where the evidence supports it.

Music represents a cornerstone of human culture, existing digitally across diverse modalities, including audio, symbolic encodings (e.g., MIDI, MusicXML), and sheet music. Despite the advancement of Multimodal Large Language Models (MLLMs), current music benchmarks face three major limitations. First, large static benchmarks are resource-intensive to evaluate, and it remains unclear how their results transfer to diverse kinds of music beyond those included in the benchmark. Second, benchmarks claiming to measure "music understanding" often fail to require music perception. Third, they do not support systematic performance comparisons across musical modalities. To overcome these issues, we introduce the Music I Care About Meta-Benchmark (MusICA-MetaBench), a framework that automatically derives on-demand benchmarks directly from user-provided data. By leveraging structured symbolic representations (e.g., MusicXML) and our pre-defined question templates, we build multiple-choice question-answer pairs that probe music perception competencies, aligned with music pedagogy, across audio, music notation images, and symbolic files. We demonstrate our framework with the ChoraleBricks dataset, and experimentally determine benchmark sizes that ensure statistically reliable model comparisons for this setup. By comparing against text-only and white-noise baselines, we show our questions do measure music perception. Ultimately, MusICA-MetaBench represents a significant advancement in the cross-modal assessment of music perception for MLLMs. By proposing a dataset-specific benchmarking paradigm, it enables efficient on-demand evaluation of music perception capabilities.

Streaming speech-to-speech language models aim to answer spoken queries directly with synthetic speech. However, standard speech and text benchmarks do not capture whether these systems behave naturally in conversations, where timing, turn-taking, prosody, interpersonal stance, language and dialect consistency, and relationship-aware appropriateness jointly shape perceived quality. We introduce SPEARBench, a benchmark for evaluating naturalness in speech-to-speech language models from question-answer interactions. SPEARBench constructs controlled dialogue prompts from the Seamless Interaction corpus, runs inference across multiple models, and evaluates generated answers using a multidimensional protocol that covers response latency, interruptions, speech quality, ASR robustness, language and dialect consistency, emotional naturalness, interpersonal stance, and explainable distributional baselines. The benchmark includes original human answers as a reference condition and reports results for several contemporary models. Results show that current models can achieve high signal-level quality and low ASR error while still differing from human conversational behavior in latency, overlap, dialect preservation, emotional adaptation, and interpersonal stance dynamics.

Large audio-language models (LALMs) frequently hallucinate by overriding acoustic evidence with language priors. While contrastive decoding (CD) offers training-free mitigation, existing methods rely on blunt perturbations like masking or noise, leaving structured audio transformations unexplored. We explore this design space by evaluating a diverse library of targeted audio perturbations and adaptively selecting the optimal negative branch for each task and example. First, we improve upon earlier prompt engineering by showing that a simple binary yes/no constraint reduces the model's tendency to falsely confirm absent audio features. Second, evaluating our library across temporal, spectral, frequency, and amplitude domains reveals that optimal transformations are highly task-dependent; for instance, reversing the audio array disrupts temporal coherence, raising accuracy on the temporal order task from 74.7% to 81.4%. Finally, we trained a light-weight perturbation selector on model hidden states to dynamically route negative branches, yielding an additional +4.3% gain on the existence task.

Recent music audio-language models achieve high accuracy on instrument question-answering benchmarks, but it remains unclear whether this reflects robust audio grounding or benchmark-specific shortcuts. In this paper, we introduce an OpenMIC-derived diagnostic benchmark sequence for instrument grounding in music audio-language models, extending binary instrument-presence QA to genre-prior-reduced examples, confusable instrument discrimination, longer audio context, and temporal localization. Across these settings, high binary QA accuracy often fails to predict model behavior: models can exhibit option-position bias, confusable-instrument errors, and temporal response bias. These results suggest that instrument grounding should be evaluated with multi-axis diagnostic benchmarks rather than a single aggregate accuracy.

Speech-capable models are increasingly deployed in real-world applications across languages. Yet their safety and fairness beyond English settings and under naturalistic conditions remain understudied. We survey safety reporting practices across state-of-the-art speech model releases, finding that only 8% document any multilingual analysis. To address this gap, we introduce RedVox, a multilingual safety and fairness benchmark for audio and speech built on real voices, covering unsafe and unfair stereotypical requests across five languages (English, French, Italian, Spanish, and German). Evaluating eight state-of-the-art models, we find that vulnerabilities persist even under non-adversarial conditions, worsen in non-English languages, and are amplified when the request comes from a spoken input. Finally, by surveying the participants who contributed to RedVox, we document the unique personal and privacy challenges of collecting speech data with human participants, pointing to broader sociotechnical challenges in naturalistic speech safety research.

This paper describes our submission to the IWSLT 2026 Instruction Following shared task. SpeechLLMs are developed for both short-form and long-form speech instruction following under constrained settings. For the short track, strong performance is achieved on MCIF, with a SIFS score of 2.0708. For the long track, three speech segmentation methods are explored, and the HIFS score is introduced to account for unstable long-form generation. Experimental results show that fixed 30-second segmentation provides the most robust long-form performance, achieving the highest HIFS score of 2.0663. Further analysis shows that hallucination mainly manifests as repetitive insertions in generated outputs, substantially affecting ASR and SSUM, while short-form capabilities are largely retained after long-form extension.

Speech conveys information through both words and vocal delivery. We evaluate four leading production realtime voice systems-OpenAI's GPT Realtime 2, Google's Gemini 3.1 Flash Live, and Alibaba's Qwen3.5 Omni Plus and Omni Flash-on tasks where the words and the delivery patterns both convey meaningful information. Across three consequential scenarios, all four systems act on the words rather than the voice. They end calls with crying callers who insist nothing is wrong, approve wire transfers authorized in frightened voices, and enroll callers whose agreement is clearly sarcastic. Surprisingly, this is often not a failure of perception. When asked directly, three of the four systems reliably identify the distress, fear, or sarcasm they later ignore when making decisions. We observe a similar pattern when these realtime voice systems estimate accent and age, as their responses frequently follow the biases of the words rather than the acoustic properties of the speaker. We term this disconnect between perception and action the emotional intelligence gap of voice AI. Prompting systems to explicitly attend to vocal delivery improves performance only partially and inconsistently. Our findings show that current realtime voice AI systems often behave as if speech had been reduced to a transcript, suggesting that they should be used with caution in settings where the tone and emotion of delivery convey important information.