Hi everyone, we’re the team behind Cue v5, a realtime decision layer for voice agents. A voice agent can have good transcription, reasoning and speech synthesis, yet still interrupt at the wrong moment or respond before someone has finished. We built Cue to handle those conversational decisions. The LLM handles what to say. Cue helps decide when to speak, when to wait and when to yield. We think of Cue as a System 1 style conversational reflex: fast decisions from incoming audio, working alongside the LLM’s deeper reasoning. Cue v5 listens to both caller audio and the agent’s own audio. It evaluates the conversation every 160 ms and supports STOP, PAUSE and RESPOND actions, helping distinguish interruptions and turn endings from backchannels, background sounds and pauses within an unfinished thought. Results from our evaluations: • TurnBench: 80% interruption recall, with no TurnBench training data. • AMI meetings: interruption recall increased from 24% with v4 to 61% with v5, at operating points with no more than 10% false alarms. The weights, code and research are public: Research: https://www.ioteverythin.com/res... Model: https://huggingface.co/IOTEveryt... GitHub: https://github.com/ioteverythin/cue Paper: https://doi.org/10.5281/zenodo.2... We’re also bringing Cue to our voice agent platform soon. Contributions are welcome, including integrations, testing across languages and accents, bug reports and runtime improvements. If you build voice agents, what causes the most trouble in your current system: interruptions, pauses or knowing when someone has finished? We’d love feedback and examples to test against.

Cue v5

Help voice agents know when to stop, pause and respond

Help voice agents know when to stop, pause and respond