Inworld open-sources Voice Isolator and SAVA voice models under Apache 2.0
realmrfakename · x · 2026-11-19
Inworld open-sourced two voice models and published the technical report behind Realtime TTS-2:
- Voice Isolator: removes background chatter, music, and noise while retaining laughs, breaths, and character; works at 48 kHz with an offline model for recordings and a realtime streaming model with 20 ms lookahead
- SAVA (Semantic-Acoustic Voice Activity): uses both sound and meaning of speech to detect speaking state and turn completion; already runs in Inworld's voice products
Models and weights are Apache 2.0, free for commercial use, available on GitHub and Hugging Face.
Related event: Inworld Open-Sources 48kHz Voice Isolator Under Apache 2.0(3 posts)→