Voice Memory Architecture for Agentic Speech Recognition

Voice Memory implements a listener-thinker architecture for ASR that uses a frozen corrector and an asynchronous, score-gated optimizer to manage a domain-specific memory file. This approach reduces error correction rates on financial news from 64% to 35% by preventing unconstrained generative over-correction.