Tool Reviews

VoxLock Extracts the Target Speaker for Voice Agents Without Requiring Enrollment

VoxLock is a 27M-parameter streaming isolator for voice agents, needs no enrollment, scores 4.95% WER on Krisp, runs on one CPU with 24 ms latency.

VoxLock Extracts the Target Speaker for Voice Agents Without Requiring Enrollment

Key Takeaways

  • VoxLock is a 27M-parameter streaming voice isolator for voice agents that can target the caller’s voice without enrollment.
  • It achieves 4.95% WER on the Krisp Voice Isolation Benchmark, reported as the lowest among tested paid and open models, with only 0.11 leaked background words per minute.
  • The model runs on one CPU core with 24 ms algorithmic latency, making it suitable for real-time conversational agents.
  • VoxLock is available as a Pipecat filter, a LiveKit Agents plugin, and in sherpa-onnx, with Apache-2.0 weights, training code, and benchmark tools released.

Meta description: VoxLock is a 27M-parameter streaming voice isolator for voice agents that needs no enrollment and scores 4.95% WER on the Krisp Voice Isolation Benchmark.

VoxLock Extracts the Target Speaker for Voice Agents Without Requiring Enrollment

Tags: voice agents, speech recognition, real-time audio, machine learning, open source

![Abstract friendly illustration of a microphone separating a clean waveform from background noise for a voice agent](voxlock-voice-isolation-illustration.png)

A new streaming voice isolator called VoxLock is aimed at teams building voice agents. The project's author says the 27M-parameter model can isolate a caller's voice from the first seconds of a call without a separate enrollment step.

VoxLock Extracts the Target Speaker for Voice Agents Without Requiring Enrollment

Why the benchmark numbers matter

On the Krisp Voice Isolation Benchmark, which uses 265 real recordings, VoxLock records 4.95% WER, or word error rate. The author says that is the lowest among every model tested, paid or open, with Krisp VI 2.5 at 5.30%. The benchmark also reports 0.11 leaked background words per minute, meaning fewer unintended background words reach the transcript or agent output.

Low latency and simple deployment

The model runs on one CPU core with 24 ms algorithmic latency. For conversational agents, that timing matters because small delays can make responses feel slow. VoxLock ships as a Pipecat filter, a LiveKit Agents plugin, and in sherpa-onnx, giving builders several integration points. The weights, training code, and the benchmark harness are released under Apache-2.0.

What readers can do next

The author notes that a browser demo is available for testing the isolator. For voice-agent teams, the practical takeaway is a no-enrollment option that reports strong benchmark results while staying small enough for CPU-only use. The Apache-2.0 code and weights make it easier to evaluate inside existing pipelines.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.