Story

In brief

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