Trelis releases voice isolation and barge detection models, free for small businesses
alexcovo_eth · x · 2026-11-19
Trelis Research released two speech models on Hugging Face: a Voice Isolator and a Barge Detection model, licensed under Polyform Small Business — free commercial use for companies under 100 employees and $1M revenue, with annual licensing for larger ones.
Design notes: the barge detector, used in Trelis's cascade, reduces false interruptions from backchannels like "mhm" and "yes"; in Turnbench comparisons it beat Silero VAD and VAP OTO on false interruptions with a small recall trade-off. The isolator suppresses background speech/noise in 16 kHz mono audio, runs on CPU with no speaker enrolment or cloud service, and is recommended before barge detection (used selectively for ASR since it can suppress speech).
More from Multimodal
- Lunch-break demo: ChatGPT + CapCut plugin turns camera roll photos into a GTA 6-style intro video — alifcoder · 2026-11-19
- NVIDIA to demo Cosmos Data Factory pipeline for training-ready synthetic datasets with Voxel51 and Encord — NVIDIAAI · 2026-11-19
- Gradium designed 16,000+ voices — here are the five parts of a good voice prompt — rdesh26 · 2026-11-19
- AI video recasts The Weeknd as Michael Jackson in striking demo — TMcFly · 2026-11-19
- AI film Gods Don't Give Gifts lands worldwide theatrical release on January 22 — DavidmComfort · 2026-11-19
- Same-seed image model shootout: Minimax H3 edges out Z-Image Turbo and Krea2 — Engine_Hungry · 2026-11-19