← All tools

Audio Enhancement

Inworld Voice Isolator

Visit website ↗

About Inworld Voice Isolator

Inworld develops Voice Isolator, an open-weight audio model for separating a foreground voice from noise, music, and other surrounding sounds. Version 1 includes an offline model for recorded material and a real-time model for live audio, both operating on 48 kHz audio and distributed under the Apache 2.0 license. The real-time version is designed for streaming, with a documented 63 ms algorithmic latency at 20 ms of look-ahead. The system preserves the foreground speaker’s recorded voice and timing, but it is a denoising tool rather than a restoration system, and it generally keeps the louder voice when people speak over one another.

For marketing and communications teams, it can clean interview recordings, improve the clarity of customer testimonials, and prepare speech datasets or voice-production pipelines. The real-time model may also help make live calls, streams, or event audio easier to hear, while the offline version can separate voice and background tracks for editing. The tools run on CPU and supported NVIDIA or Apple Silicon hardware, with GPU requirements and performance depending on the device. Teams should review speaker consent and recording rights when processing or repurposing voice recordings.

What you can use it for

  • Clean interview, testimonial, and podcast recordings
  • Improve voice clarity in live streams and calls
  • Prepare speech datasets for voice or audio workflows
  • Separate speech from music and environmental noise before editing
  • Extract a background or residual track for audio production