Inworld open-sources a voice isolation model: 48 kHz, Apache-2.0, just 58M/64M params
mrfakename0 · reddit · 2026-11-19
Inworld released a new speech isolation model that's a pleasant surprise for the OSS community, on top of its already strong TTS lineup.
- 48 kHz support, Apache-2.0 licensed
- Two variants: 64M (offline/non-streaming, 8s chunks) and 58M (realtime, 20ms chunks)
- Trained on 10k hours of audio; cheap and fast to run
- It's an isolation model, not a speech enhancement/restoration model like Resemble Enhance
The attached screenshot shows a spectrogram demo of the model processing noisy input.
Related event: Inworld Open-Sources 48kHz Voice Isolator Under Apache 2.0(3 posts)→
More from Multimodal
- Inworld open-sources 48kHz real-time voice isolation models under Apache 2.0 — Scobleizer · 2026-11-19
- ElevenLabs MCP demo builds an animated story live — images, music, and voiceover on the fly — ciguleva · 2026-11-19
- NodeUI 1.3 lets you share photos straight from your gallery into ComfyUI workflows — guhl-design-studio · 2026-11-19
- Fable 5.2 beats Astra 6.2 and Opus 5.5 at launch videos by a wide margin — haltakov · 2026-11-19
- Developer adds a real-time talking-head avatar to his local voice agent using open-weight video model — Scobleizer · 2026-11-19
- Magnific launches Illusion, an AI video editor that detects scene elements — techhalla · 2026-11-19