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

Ying Fang, Zhifu Gao, Qing Gu, Jiajun He, Kangqi Jing, Xiangang Li et al.

arXiv 2610.21614 13 pages Read the paper

The short version

The problem: Audio systems can understand sound, but when people need millions of quick, simple decisions—such as “is this the same speaker?” or “which sound is this?”—having the system write a full sentence first wastes time and computing power. The text may also need extra parsing before another program can use it.
The idea: The paper adds a direct answer interface to an existing audio model. Instead of writing a sentence, the model can choose among listed options, give the probability that a yes-or-no statement is true, or assign a rating by combining probabilities across ordered levels. It also trains these abilities together even when each audio example has labels for only some of the tasks.
What it got: The system was about 20 times faster than the original model in the tested setting. It reached 84.06% accuracy on one audio-understanding test and a 0.1062 normalized average rating error on overall speech quality.

Modules from this paper

Joint Learning from Partially Labeled Audio Tasks

Audio tasks arrive with different kinds of labels and many missing annotations; Qwen-Audio-Jev packages them in one supervision format and applies only the observed primitive’s matching loss.

audiosupervised-learningmulti-task-learningpartial-labels

Typed Decisions from Question-Conditioned Audio

Generated answer text is an indirect way to make bounded audio decisions; Jev reads the pre-answer decoder state into a typed probability or rating output.

audiomultimodaldecision-headclassification

Reinforcement Learning for Calibrated Decisions

How an optional post-supervised training stage turns sampled audio decisions and known labels into updates that target probability quality, not only correct answers.

reinforcement-learningcalibrationaudiopolicy-gradient