New Inworld voice-isolator model for speech noise removal on HuggingFace
On November 14, 2026, the model inworld/voice-isolator was published on HuggingFace with the audio-to-audio task. It isolates voice from noisy recordings and live audio at 48 kHz. The repository has two variants: voice-isolator-v1 for offline processing of recordings and voice-isolator-v1-realtime for streaming audio with 20 ms latency. The model weighs 0.5 GB, the largest file is denoiser.safetensors at 0.2 GB, weights in safetensors format. License — apache-2.0, the code is released as open source.
What it means
Our audio generators section does not yet have tools for cleaning speech from noise. The closest related ones are speech and music generators: XTTS from Coqui, YuE from M-A-P and Fish Speech from Fish Audio. All of them solve the opposite task — synthesis of speech or music, not isolation of voice from a recording.
By weight, voice-isolator is noticeably lighter: 0.5 GB versus 1.9 GB for XTTS and 11.6 GB for YuE. The apache-2.0 license of voice-isolator matches YuE, whereas XTTS is limited to non-commercial use, and Fish Speech — to revenue up to 500 thousand dollars. The author of voice-isolator does not specify VRAM requirements, unlike YuE with its stated 24 GB.
The model fills a gap in the reference: cleaning recordings before publication and preparing speech datasets is a separate class of tasks that we did not have before.How the method works. The diagram is drawn based on this news note.