NVIDIA's Voice Memory: Optimizing Speech Recognition via Editable Memory Files
nvidia · hf · 2026-07-31
NVIDIA introduced Voice Memory, an inference-only scheme for agentic speech recognition utilizing a listener-thinker architecture:
- Listener: Reads a per-domain memory.md file at stream time to decide whether to correct the current recognition hypothesis.
- Thinker: Asynchronously updates the memory file, accepting edits only if they strictly improve a held-out score.
This design requires no weight changes, keeping learned skills auditable and portable. Experiments show that unconstrained generative error correction often over-corrects, breaking correct tokens (up to 64% of the time on financial news), whereas Voice Memory reduces this rate to 35%. Across ten domains, the scheme lowers the weighted word error rate (WWER) from 8.36% to 7.52% without adding inference parameters.
More from Research
- Retriever: A Framework for Asynchronous, Closed-Loop Robot Agents — ZeYanjie · 2026-08-24
- Converting GMMs ↔ PEFs for fast KLD approximation — FrnkNlsn · 2026-08-24
- Netflix details its production LLM judge: hundreds of thousands of recommendations scored weekly — omarsar0 · 2026-08-24
- Nature Comment: Provenance, not interpretability, grounds trust in autonomous science — gabepgomes · 2026-08-24
- New Architecture RHEA: Train 1B Model on 8GB VRAM — zemondza · 2026-08-24
- Trained two 16M-param models to do generative CAD with real physics — debreuil · 2026-08-24