Voice Memory for Agentic Speech Recognition
The paper’s core claim is that a frozen ASR corrector can improve by editing a small external “memory,” without changing model weights.
Voice Memory lets a corrector decide during streaming whether to revise an ASR hypothesis or leave the 1-best output alone. A separate optimizer updates the memory only when a bounded edit improves a held-out score. The authors report weighted word error rate dropping from 8.36% to 7.52% across ten HyPoradise domains, with no dataset falling below its baseline. They frame the main gain as restraint: fewer harmful over-corrections than unconstrained generative error correction. HF Daily Papers' note
Voice Memory lets a corrector decide during streaming whether to revise an ASR hypothesis or leave the 1-best output alone. A separate optimizer updates the memory only when a bounded edit improves a held-out score. The authors report weighted word error rate dropping from 8.36% to 7.52% across ten HyPoradise domains, with no dataset falling below its baseline. They frame the main gain as restraint: fewer harmful over-corrections than unconstrained generative error correction. HF Daily Papers' note
score 4