Qwen-Audio-Jev: A Unified Interface for Audio Decisions

Peixin Chen; Zhigao Chen; Zhiyun Fan; Ying Fang; Zhifu Gao; Qing Gu; Jiajun He; Kangqi Jing

cs.SDOct 15, 2026

Qwen-Audio-Jev: A Unified Interface for Audio Decisions

Authors: Peixin Chen, Zhigao Chen, Zhiyun Fan, Ying Fang, Zhifu Gao, Qing Gu, Jiajun He, Kangqi Jing,

Organizations: Token Foundry, Alibaba Group

Abstract

Large audio language models (LALMs) have made substantial progress in audio understanding. However, autoregressive text generation introduces computational overhead and latency in high-throughput applications that require frequent selection, judgment, and scoring. To address this limitation, we introduce Qwen-Audio-Jev, which augments Qwen-Audio-3.1-ASR-Flash-Next with a dedicated Jev decision head. The model unifies selection, binary judgment, and scoring through three decision primitives: CHOICE, NOUL, and SCORE. Given an audio input, a natural-language question, and an explicit answer specification, Qwen-Audio-Jev directly outputs a task-conditioned probability distribution or a numerical score through typed decision readouts. Evaluation across 18 benchmarks demonstrates the model's effectiveness: CHOICE achieves 78.40 percent accuracy on MMAU and 83.14 percent on MMSU; NOUL achieves 93.76 percent accuracy on the VoxCeleb1-O speaker verification benchmark; and SCORE achieves a normalized mean absolute error of 0.1235 on NISQA overall quality assessment. These results highlight the potential of Qwen-Audio-Jev for efficient audio decision-making in applications such as large-scale audio screening, speaker verification, and automated quality assessment.

Figures & tables

Figure 1: Qwen-Audio-Jev model architecture. Audio and serialized question/criteria inputs feed the shared Qwen-Audio-3.1-ASR-Flash-Next decoder. Its final-prefix hidden state, computed before target-answer tokens, supports three typed affine readouts. Choice and Score normalize only valid candidate or level positions. The requested primitive selects the active output; the three readouts do not imply simultaneous multi-question inference.

Type	Answer space	Native output
Choice	2–64 ordered candidate descriptions	Candidate probabilities and the selected option
Noul	A binary proposition	Probability that the proposition is true
Score	2–10 ordered rating levels	Level probabilities and the expected level index

Table 1: Answer spaces and native outputs of the three decision readouts in the current implementation.

Figure 2: Training pathways of Qwen-Audio-Jev. (a) SFT compares predicted distributions with observed labels and directly backpropagates task-specific losses. (b) The optional RLCD-inspired stage initializes from SFT, samples multiple decisions, and uses label-based group rewards for reward-weighted policy-gradient updates. Choice, Noul, and Score are alternative question-conditioned tasks; only the requested primitive is active.

Benchmark	Accuracy (%) $\uparrow$
CochlScene	78.9645
NSynth	74.2432
NonSpeech7k	90.4828
DiDiSpeech	100.0000
MELD	58.1511
MMAU	78.4000

Table 2: Choice accuracy (%) of Qwen-Audio-Jev on eight labeled benchmarks covering acoustic-scene classification, musical-instrument recognition, human non-speech sound recognition, speaker-gender classification, speech emotion recognition, and audio question answering.

Configuration	Accuracy (%) $\uparrow$	EER (%) $\downarrow$	AUC $\uparrow$	Brier $\downarrow$	FPR (%) $\downarrow$	FNR (%) $\downarrow$
VoxCeleb1-O	93.7595	6.1714	0.983857	0.047376	5.6199	6.8613
Clotho-AQA	79.7802	19.0659	0.884508	0.148500	12.0101	26.5306
PodSarc	99.6877	0.3195	0.999898	0.003059	0.3195	0.3037
In-the-Wild	74.0072	22.8112	0.850632	0.167138	3.6870	63.6764

Table 3: Noul performance of Qwen-Audio-Jev on speaker verification, binary audio question answering, sarcasm detection, and audio-deepfake detection. Accuracy, FPR, and FNR use a probability threshold of 0.5.

Configuration	MOS MAE $\downarrow$	MOS RMSE $\downarrow$	Normalized MAE $\downarrow$	Normalized RMSE $\downarrow$	Pearson $\uparrow$	Spearman $\uparrow$
NISQA overall	0.493912	0.615819	0.123478	0.153955	0.779906	0.753822
NISQA noise	0.557110	0.720885	0.139278	0.180221	0.654883	0.688694
NISQA coloration	0.410769	0.514844	0.102692	0.128711	0.809775	0.820855
NISQA discontinuity	0.419045	0.511465	0.104761	0.127866	0.860221	0.825081
NISQA loudness	0.453387	0.564739	0.113347	0.141185	0.676132	0.608626
VoiceMOS	0.371024	0.466741	0.092756	0.116685	0.880209	0.872587

Table 4: Score performance of Qwen-Audio-Jev on NISQA speech-quality dimensions and VoiceMOS/BVCC synthetic speech. Error and correlation metrics summarize numerical accuracy and agreement with human quality ratings.

Audio-Language Models

Audio-language models connect audio and natural language for understanding, retrieval, dialogue, speech generation, and music or sound analysis.

See all 303 papers

AudioJev: Direct Audio Decisions with Order-Calibrated Probabilities
Oct 1, 2026Sihan Lv, Zhen Li, Zhiqi Cao +4Audio QA Audio Reasoning
Zhejiang University, Hangzhou, China · China Academy of Space Technology, Beijing, China · Innovation and Management Center of the School of Software (Ningbo), Zhejiang University, Ningbo, China +2
AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models
Oct 19, 2026Marwan Farag, Bin YangToken Pruning Audio-Language Models
Institute of Signal Processing and System Theory (ISS), University of Stuttgart, Stuttgart, Germany
AS $^2$ D: Accelerating On-Demand Audio Understanding on Mobile Devices
Sep 29, 2026Yunzhe Li, Kyoungjun Park, Hongzi Zhu +1On-Device Language Model Inference Language Model Decoding
The University of Texas at Austin · Shanghai Jiao Tong University

Oct 1, 2026J/K move · Enter open · S save

AudioJev: Direct Audio Decisions with Order-Calibrated Probabilities

Sihan Lv, Zhen Li, Zhiqi Cao, Jinshan Zhang+3

Zhejiang University, Hangzhou, China · China Academy of Space Technology, Beijing, China · Innovation and Management Center of the School of Software (Ningbo), Zhejiang University, Ningbo, China+2

Audio QA Audio Reasoning Audio-Language Models Classification Permutation-Invariant Learning

Open paper AskPDF arXiv

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.

Abstract

Audio-Language Models

Explore similar work

AudioJev: Direct Audio Decisions with Order-Calibrated Probabilities

AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models

AS2^22D: Accelerating On-Demand Audio Understanding on Mobile Devices

AS $^2$ D: Accelerating On-Demand Audio Understanding on Mobile Devices