Reduce first-packet latency without sacrificing accuracy

Reduce the first-output startup delay of VibeVoice-ASR-Streaming below the current full-chunk-plus-lookahead latency while preserving the accuracy benefits of the larger 22-frame chunks.

Background

VibeVoice-ASR-Streaming reports a steady-state speaker-attribution latency of 2.00 seconds for the 22-frame configuration and 1.53 seconds for the 15-frame configuration. However, the first output incurs an additional startup delay because the system must receive one complete chunk together with its lookahead: 3.5 seconds at 22 frames and 2.5 seconds at 15 frames.

The paper identifies reducing this initial delay while retaining the recognition and speaker-attribution accuracy associated with larger chunks as unresolved future work. This is a concrete deployment problem for real-time voice assistants and agents, where startup latency affects responsiveness even when subsequent outputs meet the target streaming latency.