DIFFA-2: A Practical Diffusion Large Language Model for General Audio Understanding

Benchmark Model Rank Results
audio-understanding-reasoning-on-mmarDIFFA-2 (8B)–Accuracy: 50.8
audio-understanding-reasoning-on-mmau-testDIFFA-2 (8B)–Accuracy: 67.0
audio-understanding-reasoning-on-mmsuDIFFA-2 (8B)–Accuracy: 60.45
speech-recognition-on-librispeech-test-otherDIFFA-2 (S1)–Word Error Rate (WER): 5.34