Cue v5

Cue v5 acts as a real-time decision layer for enhancing voice agent interactions. Key capabilities include: * Real-time audio analysis for caller and agent * Precise control over agent responses (STOP, PAUSE, RESPOND) * Distinguishes interruptions from backchannels * Identifies completed turns versus thinking pauses * Compact and causal conversation control model This system continuously listens to both caller and agent audio streams, providing decision points every 160 ms. It significantly improves the flow of conversations by intelligently managing when a voice agent should yield, pause, or deliver a response. Unlike traditional voice-activity detectors or silence timers, Cue v5 understands the nuances of human conversation, preventing agents from interrupting mid-turn pauses or responding prematurely during backchannels. The latest version, v5, integrates optional features from the agent's own audio, building upon its audio-only predecessors while maintaining robust leakage-aware training. It is trained on a diverse mixture of synthetic and real conversational data, including Indian-English recruiter calls and genuine meeting recordings, ensuring adaptability to natural conversation patterns. This advancement addresses the synthetic-to-real gap, improving performance in authentic dialogue settings. Ideal for contact centers, virtual assistants, and conversational interface developers, Cue v5 enables more natural, efficient, and user-friendly voice agent interactions. It provides public weights, code, and research, encouraging contributions from the developer community.