⚡️Alibaba introduced Qwen-Audio-Jev, adding a dedicated Jev decision head to Qwen-Audio-3.1-ASR-Flash-Next for high-throughput audio decision tasks.
Instead of generating text answers, it directly outputs choices, binary judgments, or scores through three primitives: CHOICE, NOUL, and SCORE.
Inference speed is ~20× faster than the original model.
Evaluation across 18 labeled configurations covers all three primitives:
• 84.06% accuracy on MMSU
• 6.50% EER on VoxCeleb1-O speaker verification
• 0.1062 normalized MAE on NISQA audio quality assessment
显示更多