AnalysisAI ModelsOctober 30, 2026

VoxLock extracts target speaker for voice agents without enrollment

Read original source →arxiv.org

A fine-tuned causal target-speaker extraction model reaches near-commercial speech recognition accuracy on real recordings while keeping CPU usage and latency low. The 8-bit model needs no enrollment step, and the project has 85 GitHub stars.

1 source

More stories today

Open the live feed