The paper investigates how to evaluate audio‑language models by separating the use of acoustic evidence from the need to invoke a generative audio model. Using a controlled call‑decision framework, the authors compare policies that rely on transcript labels, encoder outputs from CLAP, AST, or WavLM, and optional calls to generative models such as Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑only controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
By Mengzhe Geng
The paper introduces a reproducible framework that converts existing text-based tool‑calling benchmarks into audio‑based evaluations using text‑to‑speech, speaker variation, and environmental noise, while preserving original annotations. Experiments on seven omni‑modal models show that performance varies strongly by model and task, with notable gaps between text and voice results. The study also includes a stress test, a reference‑free LLM‑as‑judge protocol, and demonstrates that open‑source judges can match proprietary ones, aiding privacy‑preserving evaluation.
By Md Tahmid Rahman Laskar, Xue-Yong Fu, Seyyed Saeed Sarfjoo, Quinten McNamara, Jonas Robertson, Shashi Bhushan TN
The study evaluates various uncertainty estimation methods—probability-based, sampling-based, self-verification, evidential, and contrastive—across four open-weight audio-language models and five audio QA benchmarks. In multiple-choice settings, first-token probability measures outperform others, achieving a mean AUROC of .740, while open-ended evaluation shows lower accuracy but still predictive uncertainty. Ablation experiments reveal that removing audio evidence significantly degrades error-detection performance, indicating that uncertainty relies more on audio than on question text.
By Aaron Isidore Grace, Weiran Wang
The paper investigates how to evaluate generative audio large language models (Audio‑LLMs) on known closed‑set tasks by separating the decision to call a generative model from the use of acoustic evidence. It introduces a controlled call‑decision framework where a policy can choose between a transcript label, encoder evidence from CLAP, AST, or WavLM, or a generative call to Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio, and measures the impact of generative calls on accuracy. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑based controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
By Mengzhe Geng
The paper introduces a symbiotic architecture that equips large language models with audio‑understanding abilities without fine‑tuning their weights. It uses an injector module to write audio‑conditioned vectors into the LLM’s key‑value cache, allowing the model to act as an audio language model while keeping the backbone unchanged. The approach improves scalability—since injection cost depends on the injector width—and preserves the LLM’s original text performance, outperforming conventional frozen‑LLM methods and approaching fine‑tuned ALM results on audio tasks.
By Yotaro Kubo, Qi Sun, Yujin Tang
arXiv:2609.08899v2 Announce Type: replace-cross
Abstract: Speech deepfakes can mimic a speaker's voice convincingly enough to deceive listeners and automated systems. This has driven strong progress...
By Mengzhe Geng, Yujia Lu, Patrick Littell, Manuela Kunz, Xie Chen
Mizar is a 159.3‑million‑parameter audio‑language model designed for devices with limited memory and computation. It couples a compact CED‑Small audio encoder with SmolLM2‑135M via a frequency‑merging mapper and is trained in three stages—audio‑language alignment, audio‑dependent fine‑tuning, and post‑training—to improve performance on audio‑question tasks. Across five random seeds, Mizar outperforms all other sub‑200M‑parameter ALMs on MMAU, MMAR, and ADQA‑clean, achieving mean accuracies of 52.92%, 42.42%, and 36.02% respectively, while enabling local inference on a single CPU with an average latency of 1.09 seconds for MMAU questions.
By Kaiyang Li, Shaobo Han, Yue Tian, Shihao Ji
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.
The Qwen-Audio-3.0-ASR Technical Report introduces a Mixture-of-Experts large language model-based automatic speech recognition system that addresses real‑world production challenges such as regional dialects, dynamic entities, hotwords, long‑range context, and disfluent speech. Built on the Qwen backbone and trained on tens of millions of hours of speech data, it supports transcription in 30 languages and 16 Chinese dialects, and offers industry‑domain entity recognition, hierarchical hotword customization, single‑pass polishing, and long‑audio contextual modeling. A streaming variant, Qwen-Audio-3.0-ASR-Streaming, is also presented for low‑latency applications, with evaluations showing state‑of‑the‑art performance against leading commercial systems.
By Chuanmeng Bian, Daren Chen, Peixin Chen, Zhigao Chen, Zhiyun Fan, Zhifu Gao, Bo Gong, Qing Gu, Jiajun He, Yawei Hu, Yunjie Ji, Jingbei Li, Xiangang Li, Xu Li, Zengxi Li, Zheng Li, Chengdong Liang, Baiji Liu, Ying Liu, Bin Ma, Yiping Peng, Yuezhang Peng, Zhendong Peng, Yu Pu, Yang Shi, Xin Shu, Jian Tang, Biao Tian, Peiyao Wang, Tianzi Wang, Wen Wang, Wupeng Wang, Cheng Wen, Yuzhong Wu, Zijian Xia, Yunchong Xiao, Nan Yang, Jianwei Yu, Jixing Yu, Binbin Zhang, Lei Zhang, Sitong Zhao, Guangdong Zhou, Yuan Zhou, Jianheng Zhuo
The paper introduces the Speech-Unsupported Rejection Evaluation Challenge (SURE‑Challenge), a benchmark designed to test whether speech‑LLMs should accept or reject audio inputs before generating answers. Using LibriSpeech‑derived transcriptions paired with first‑word question answering, the authors evaluate various noise and silence conditions, and compare a simple energy‑plus‑Whisper‑score rule against a Qwen2‑Audio front‑end. On a 474‑row test set, the rule rejects 196 of 204 unsupported inputs while preserving accuracy on supported data, revealing a pre‑generation error mode that answer‑only scoring misses.
By Mengzhe Geng
The paper presents a new evaluation protocol for post‑training quantization of speech language models that separates lexical output, transcript‑insufficient endpoints, and packed implementations. In a Qwen2‑Audio case study, a 6‑bit allocation selected for translation improves chrF scores but degrades emotion recognition, while uniform and front‑layer controls perform better on emotion tasks. Similar patterns hold at 7 bits, and a 4‑bit study shows consistent emotion deficits across all low‑bit allocations, with no advantage for the selected scheme. The study highlights a precision‑dependent mismatch between lexical output, waveform‑dependent behavior, and nominal precision, without claiming a general failure of low‑bit models or a deployment benefit for the selected allocation.
By Mengzhe Geng, Jinxi Jin, Junhao Xu
The paper introduces TORUS, a closed‑loop benchmark designed to evaluate unified audio models on their ability to understand the audio they generate or edit. TORUS contains 1,053 questions across 21 tasks and 8 audio domain settings, covering all three audio types. The authors evaluate both unified models and cascades of specialized models, finding that the best unified models answer only about 52.4% of questions about their own outputs, while GPT‑Audio achieves 53.6%.
By Aryan Vijay Bhosale, Harshit Rajgarhia, Abhishek Mukherji, Dinesh Manocha