Add inworld/voice-isolator as a noise suppression option

(#4940) Feature Open: Under consideration desktop voice

Current problem

The existing noise suppression models (RNNoise, GTCRN, DeepFilterNet) may not meet everyone's needs, and as reported in #3763, background noise such as typing and fans still leaks through with all of them. More choice is always welcome here, since different mics, rooms and CPUs need different trade-offs.

Proposed change

Please consider adding inworld/voice-isolator as another option in the Noise suppression dropdown. This is a very new model. It was released by Inworld on 18 Nov 2026 as their first open-source model, so there is little independent testing or long-term real-world feedback yet. It is already getting a lot of attention: as of 20 Nov 2026 it has 300+ likes on Hugging Face and ranks #2 on the Models page sorted by Trending. About the model
  • Open-source voice isolation model released by Inworld in Nov 2026, Apache-2.0 for both code and weights.
  • It removes background noise, music and other voices while keeping the speaker's voice (timbre, breaths, level) as recorded.
  • Full-band 48 kHz, trained on ~10,000 hours of multilingual speech mixed with many kinds of noise.
  • v1 does denoising only. It is not a speaker selector and does not restore lost audio quality.
On Inworld's own VIBE benchmark, the realtime model scores 0.718 and the offline model 0.768, versus 0.510 for DeepFilterNet3. Links

Platform

macOS, Windows, Linux, Web

Comments

Sign in with Fluxer to comment and vote.

No comments yet.