VoxLock: Target-Speaker Extraction for Voice Agents
October 29, 2026
VoxLock is a fine-tuned 27M parameter causal model that enables target-speaker extraction from a single voice sample provided within a call. It achieves a 4.95% WER on the Krisp benchmark, significantly outperforming unprocessed audio and competing with top commercial filters.
HOW THIS AFFECTS YOU
●
builderYou can implement highly effective speaker isolation in voice agents without an enrollment step.
●
founderThis offers a competitive technical path for building privacy-focused or high-clarity voice assistants.