# New voice isolation model from Inworld — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/d644b319bb2ec185

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-18 21:01 ingested 2026-11-19 02:39

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xzqpzu/new_voice_isolation_model_from_inworld/)

## retained text



```
[Video credit: Victor Mustar on X](https://x.com/victormustar/status/2123151425375187260)

New speech isolation model from Inworld, looks quite exciting. Inworld's TTS models are quite good already and this is a nice surprise for the OSS community IMO!

* Supports 48 kHz
* Apache-2.0 licensed
* Two variants (64M/58M) for offline/non-streaming (8s chunks) & realtime (20ms chunks)
* Trained on \~10k hours of audio

It is not a speech enhancement/restoration model (unlike Resemble Enhance), simply an isolation model. But still a nice release as it is cheap & fast to run.

Model: [https://huggingface.co/inworld/voice-isolator](https://huggingface.co/inworld/voice-isolator)

HF space: [https://huggingface.co/spaces/mrfakename/voice-isolator](https://huggingface.co/spaces/mrfakename/voice-isolator)
```



document fetched 2026-11-19 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzqpzu/new_voice_isolation_model_from_inworld
