Voice Memory cuts ASR errors by 10% without retraining via novel agentic loop

Voice Memory presents a new paradigm for speech recognition that separates the recognizer from a 'thinker' — an asynchronous optimizer that revises the transcription memory through bounded edits. Critically, the corrector is frozen (no weight updates), and the optimizer only accepts an edit when it strictly improves a held-out score. This prevents the over-correction common in generative error correction (GER), where up to 64% of edits break correct tokens on financial news. Voice Memory reduces that rate to 35%.

The results speak for themselves: across ten HyPoradise domains with an open corrector, weighted word error rate falls from 8.36% to 7.52% (7.47% with three in-context examples). No dataset regresses below its 1-best baseline. Gains concentrate where recoverable headroom is largest, including air-travel commands (8.40% to 3.40%) and noisy far-field speech (CHiME-4, 12.69% to 10.46%). The memory transfers across corrector families and adds zero parameters to the inference path, making it a practical, auditable, and portable improvement for production ASR systems.