Qwen-Audio-Jev
A unified audio decision interface model introduced by the Qwen team. Built on Qwen-Audio, it adds a Jev decision head that directly outputs structured decision results through three decision primitives: CHOICE, NOUL, and SCORE. It bypasses the autoregressive text generation stage, producing decisions in a single forward pass rather than through step-by-step token generation.
Timeline (last 90 days)
Qwen-Audio-Jev is implemented by adding a dedicated Jev decision head on top of Qwen-Audio-3.1-ASR-Flash-Next, turning audio decision-making from 'generate text then parse' into 'directly output the decision'.
Qwen-Audio-Jev converges the decision task into three primitives: CHOICE (select from candidates), NOUL (binary judgment), and SCORE (numerical scoring).
Qwen-Audio-Jev was evaluated on 18 groups of annotation configurations, covering all three primitives.
Qwen-Audio-Jev achieves a CHOICE accuracy of 78.70% on MMAU-mini.
Qwen-Audio-Jev achieves a CHOICE accuracy of 84.06% on MMSU.
Qwen-Audio-Jev achieves an equal error rate (EER) of 6.50% on the VoxCeleb1-O speaker verification benchmark.
Qwen-Audio-Jev achieves a normalized mean absolute error of 0.1062, a Pearson correlation coefficient of 0.8210, and a Spearman correlation coefficient of 0.8074 on NISQA overall quality assessment.
Qwen-Audio-Jev achieves approximately a 20-fold inference speedup compared with the original Qwen-Audio-3.1-ASR-Flash-Next.
The speedup of Qwen-Audio-Jev comes from eliminating the overhead of autoregressive text generation, not from a smaller model or reduced precision.
The paper explicitly points to three types of application scenarios: large-scale audio screening, speaker verification, and automated quality assessment.
3 more timeline events
All Facts (13)
Qwen-Audio-Jev is implemented by adding a dedicated Jev decision head on top of Qwen-Audio-3.1-ASR-Flash-Next, turning audio decision-making from 'generate text then parse' into 'directly output the decision'.
50%UnverifiedQwen-Audio-Jev converges the decision task into three primitives: CHOICE (select from candidates), NOUL (binary judgment), and SCORE (numerical scoring).
50%UnverifiedQwen-Audio-Jev was evaluated on 18 groups of annotation configurations, covering all three primitives.
50%UnverifiedQwen-Audio-Jev achieves a CHOICE accuracy of 78.70% on MMAU-mini.
50%UnverifiedQwen-Audio-Jev achieves a CHOICE accuracy of 84.06% on MMSU.
50%UnverifiedQwen-Audio-Jev achieves an equal error rate (EER) of 6.50% on the VoxCeleb1-O speaker verification benchmark.
50%UnverifiedQwen-Audio-Jev achieves a normalized mean absolute error of 0.1062, a Pearson correlation coefficient of 0.8210, and a Spearman correlation coefficient of 0.8074 on NISQA overall quality assessment.
50%UnverifiedQwen-Audio-Jev achieves approximately a 20-fold inference speedup compared with the original Qwen-Audio-3.1-ASR-Flash-Next.
50%UnverifiedThe speedup of Qwen-Audio-Jev comes from eliminating the overhead of autoregressive text generation, not from a smaller model or reduced precision.
50%UnverifiedThe paper explicitly points to three types of application scenarios: large-scale audio screening, speaker verification, and automated quality assessment.
50%UnverifiedThe premise of the decision-head mode is that the answer specification is sufficiently clear; for open-ended questions that require generating explanations or reasoning processes, text generation is still necessary.
50%UnverifiedThe evaluation of Qwen-Audio-Jev is concentrated on English benchmarks, and its performance in multilingual scenarios remains to be verified.
50%UnverifiedQwen-Audio-Jev adopts a 'separation of understanding and decision-making' design, so that the backbone model does not need retraining and only a small number of parameters need to be added at the output end to support new decision types.
50%