Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents
solyarisoftware · x · 2026-10-18
NineNineSix released TYGER on Hugging Face, an open-weight speech decision model for real-time full-duplex voice agents.
- What it does: listens to both the user's mic and the agent's TTS playback simultaneously, deciding every 80ms whether to speak, yield, or backchannel with an "mm-hm" — replacing traditional VAD.
- Architecture: NVIDIA Nemotron 3.5 streaming ASR (0.6B, 638M params) unchanged as the encoder, using its layer-15 features fed into a 3-layer causal GRU (only 14M params) for turn-taking; 652M params total, 2.61GB in FP32, just 9KiB of state per conversation.
- Outputs: per-frame turn state (listen/hold/speak/backchannel) plus endpoint, barge-in and backchannel scores with configurable thresholds (endpoint default 0.91).
- Ecosystem: the companion tyger-s2s voice agent (Apache-2.0) offers a browser demo and Docker image usable as a voice backend for any LLM + TTS stack; inference code lives in its runtime, not the weights repo.
- English-only for now, weights under the openmdw-1.1 license.
More from coding & agent
- Claude-written test harness let 31KB fake archives land on a real backup server — Grimmoner · 2026-10-18
- Matt Pocock's pre-commit hook caps agent-written files at 5,000 tokens — mattpocockuk · 2026-10-18
- Backups silently failed for 10 days: three monitoring blind spots lined up — Grimmoner · 2026-10-18
- Author Laurence Moroney: write your ugly first draft yourself, then let an LLM edit it — lmoroney · 2026-10-18
- Devs: Opus 5.5 on medium beats Opus on max — eyishazyer · 2026-10-18
- He Stopped Telling Claude Code to Stop Making Up Timestamps—and Made It Impossible with a PreToolUse Hook — Grimmoner · 2026-10-18