Speech-Hands: A Self-Reflection Voice Agentic Approach to Speech Recognition and Audio Reasoning with Omni Perception

In three linesSpeech-Hands is a voice-agentic framework learning when to trust its predictions versus consulting external audio perception. The model reduces WER by 12.1% across 7 OpenASR benchmarks and achieves 77.37% accuracy on audio QA, using a self-reflection mechanism to avoid noisy hypotheses.

Summary generated by Claude — human-verified