VoxLock releases open 27M-parameter model cutting voice agent WER to 4.95%
Tool · Hugging Face · stat: 4.95% WER VoxLock released an open-source 27-million-parameter target-speaker extraction model that isolates a caller's voice using audio from the live call itself.…
Tool · Hugging Face · stat: 4.95% WER
VoxLock released an open-source 27-million-parameter target-speaker extraction model that isolates a caller's voice using audio from the live call itself. Operating at 24 ms algorithmic latency on 77% of a single CPU thread, the 8-bit model cuts word error rate from 42.95% to 10.63% on AMI meeting benchmarks.
Local CPU speaker isolation makes real-time voice agents cheaper to deploy Voice AI developers can eliminate background audio interference without routing call streams through expensive proprietary filtering APIs.
Every claim ties to a primary source. See our methodology.