Fill Talker: ConvFill's 0.6B talker as a drop-in layer for open voice agents

I found no packaged Pipecat processor or LiveKit Agents plugin that runs a ConvFill-style trained talker locally and feeds the reasoner's notes back into its context. Today you can get filler speech from Inworld's closed realtime stack, from DIY LiveKit demos, or from Pipecat's general recommendation to send simple turns to a smaller model. So an open-source drop-in is still available, and the LiveKit forum thread shows people want it. The plugin itself is easy to copy, though. The weights are public, LiveKit or Pipecat could ship an official version, and Inworld already has a hosted one. The business model is also weak: indie developers who run a 0.6B model through llama.cpp on their own machines have little reason to pay for hosting it. The better paid angle is probably a hosted, low-latency talker endpoint co-located with telephony, or talkers fine-tuned per vertical for clinics and restaurants, where the training data is the moat. Filler Audit is thinner. The ConvFill paper already defines the NLI non-contradiction metric, and Cekura and GIGA already sell contradiction and hallucination checks to the same pre-launch buyers. It works better as a free feature that brings people to the plugin than as a product of its own.

• ConvFill / Thinking While Speaking (vysri)

  This is the source of the idea. The authors released talker weights from 135M to 1.7B, including Qwen3-0.6B, plus a dataset and code (the paper names the GitHub repo vysri/conversational-infill, though my search didn't find it directly). Their eval already has a Non-Contradiction metric, scored with an NLI classifier, that checks whether the talker's filler contradicts the reasoner. That is close to what Filler Audit would do. I found no Pipecat or LiveKit packaging from the authors. Paper: https://arxiv.org/pdf/2511.07397

• Inworld Realtime API responsiveness layer

  This is a hosted feature, not a plugin. A small filler LLM runs alongside the main model and speaks if the main model's first token is late, with settings for the filler model and deadline. It shows that a commercial vendor already sells this pattern, but inside its own closed stack.

• webrtc.ventures LiveKit fast pre-response project

  A consultancy write-up of a LiveKit agent where a small model gives an immediate reply while a larger model writes the full answer, run in parallel. It's DIY sample code, not a maintained plugin, and it doesn't feed the reasoner's notes back into the talker.

• LiveKit community: fast vs reasoning model switching

  Developers are asking for this. The reply in the thread says the only way to swap models mid-call is to swap the whole Agent, so LiveKit has no built-in talker/reasoner split.

• Cekura

  A funded voice-agent QA startup that runs simulated pre-launch calls and checks whether spoken answers are grounded in the knowledge base. It isn't specific to filler, but it targets the same pre-launch buyer as Filler Audit and could add a filler check easily.

• GIGA hallucination detection

  A vendor detector that runs during the speech gap and separates contradictions from fabrications in production voice agents. It overlaps with the per-call contradiction scoring idea.

• Talker-Reasoner (Google DeepMind paper)

  An earlier research version of the fast-talker, slow-reasoner design. It's prior art only, not a product.

Checked 2026-10-27