saolabsai/cue is a tool that helps voice agents judge in real time whether the other party's utterance has ended. According to the repository description, a Rust audio processor runs on the CPU, takes continuous audio as input, and emits decisions related to end-of-turn. It does not handle speech recognition or answer generation. When to respond after receiving a decision is up to the application that uses it. The repository does not list a separate latest release version.
Cue combines the pretrained Smart Turn v3.2 with Silero VAD. The results are split into four categories — end, hold, backchannel and interrupt — and are used to decide whether the agent should respond right away or keep listening. Backchannel means a brief interjection in the middle of a conversation. Python and Pipecat integration are optional, and Pipecat is a framework for connecting voice agent pipelines. The repository's primary language is Rust and its license is MIT. Cue's role, therefore, is not to recognize conversation content or produce answers, but to pass along utterance state so that a response time can be set.
The README notes that it has not been demonstrated that Cue is more accurate than Smart Turn. The default evaluation was conducted with synthetic material, and a performance advantage in real conversations is not an established fact. The four-way decision can therefore be regarded as a capability the agent can use, but how stably it works in real conversations is hard to judge from this material alone. For now, it is more appropriate to view it as a component for end-of-turn detection than as an alternative with validated performance.