Inworld open-sources 48kHz real-time voice isolation models under Apache 2.0
Scobleizer · x · 2026-11-19
Inworld released Voice Isolator v1 on Hugging Face under Apache 2.0, with two models sharing one architecture: an offline 64M-param model for recordings and dataset curation, and a 58M-param causal streaming model with 20ms look-ahead for live calls and streams.
Key points:
- 48kHz full-band voice isolation targeting noisy recordings and live audio
- Offline model sees full 8s chunks per frame for best quality
- Subtracting the isolated voice extracts background noise or backing-track residual for vocals/music separation
- Ships as a pip package with AGENTS.md so coding agents can install and run the first denoising pass
- Tester victormustar calls it impressive and real-time capable; Scobleizer compares it to the tape hiss of the 1970s and calls the result stunning
More from coding & agent
- GPT5.5 ports DSPy-style axllm to Zig in 45 minutes for $16 — dosco · 2026-11-19
- Pichai Spotlights Photorealistic 3D Lab Modeled in Antigravity via Blender MCP — sundarpichai · 2026-11-19
- Google's Gemini 4 Neon goes GA in GitHub Copilot, tuned for agentic coding — intellectronica · 2026-11-19
- Dev shares Gemini 4 Neon remote coding interface screenshot — cgarciae88 · 2026-11-19
- Ex-bootcamp teacher: frontier models code like talented bootcamp students — dansitu · 2026-11-19
- Blog Argues Labs Should Train RL Over a Distribution of Human Effort Levels — voooooogel · 2026-11-19