Turn Kit: drop-in full-duplex turn-taking for open-source voice agents

I could find nothing public about Tyger itself, so this does not compare against its actual accuracy. Swapping in a better end-of-turn or interruption detector for Pipecat and LiveKit is already covered, mostly for free, by Smart Turn, LiveKit's own models, Krisp, Deepgram Flux and open models like voxturn. A plugin that only does that would struggle to get anyone to pay. The real opening is the agent chiming in on its own. Every existing tool focuses on whether the user has finished and whether a sound is a real interruption or just an uh-huh. None I found lets the agent decide to backchannel, ask a clarifying question or politely cut in while the user is still talking, with one setting for how often. Research is describing this kind of model, but no framework ships one yet. A second opening is self-hosting: LiveKit's interruption model only runs on its own cloud, so a hosted or local model that does the same job plus chime-ins and works in both frameworks has a hook. For Barge-in Bench, TurnBench already covers scoring models on a fixed dataset and Cekura covers simulated end-to-end tests. The remaining gap is a simple tool that replays an agency's own clinic or tradesperson calls, with phone-line noise and background sound, through several detectors side by side. That works best as a free tool that brings people to the paid model rather than as a product on its own. I did not confirm whether Cekura can replay real recordings, so that should be checked. The main risk is that the competition is free and built into the frameworks, so the paid model needs clearly better numbers on the published benchmarks.

• Pipecat Smart Turn (Daily)

  Open model under a permissive BSD licence that detects from the audio when the user has finished. It ships bundled with Pipecat and has been the default way Pipecat ends the user's turn since v0.0.102. Covers 23 languages and runs on CPU in roughly 10 to 100 ms. Weights, training script and data are all public.

• LiveKit turn detector and adaptive interruption handling

  LiveKit Agents turns on its open-weights turn detector by default and adds an adaptive interruption model that tells real barge-ins from uh-huh style backchannels. The interruption model only runs on LiveKit's hosted service, with limited use for local development, and forum users are asking for a self-hosted version.

• Krisp VIVA turn prediction and interruption prediction

  Commercial models that work on audio alone to predict end of turn and separate real interruptions from backchannels. Pipecat has ready-made integrations and Krisp also supports LiveKit.

• voxturn

  Open turn-taking model that publishes its own false cut-off rates at a 300 ms latency budget against Baton, Deepgram Flux, Smart Turn and LiveKit. A sign that more open alternatives keep appearing.

• Sesame TurnBench

  Public benchmark and leaderboard scoring 14 systems on end-of-turn recall, interruption false positives on backchannels, noise and echo, and latency. Built on 30 hours of hand-labelled two-channel conversations, with a 104-hour training set released. Overlaps heavily with Barge-in Bench, but uses Sesame's own recordings rather than the customer's calls.

• Cekura

  Funded voice-agent testing startup ($2.4M raised) that tests barge-in from start to finish: it detects talk-over during the agent's speech, measures how long the agent keeps talking and whether the call recovers, and benchmarks Retell, Vapi, Pipecat and LiveKit. I could not confirm whether it replays real recorded calls.

• Gradium semantic VAD

  Vendor selling meaning-based end-of-turn detection as an upgrade on silence-based detection. It says itself that this is not full duplex.

Checked 2026-10-17