Hugging Face has published a technical report on StepAudio 3 Realtime, an audio-language foundation model designed for real-time spoken interaction. The system is organized around a continuous listen-converse-think-act loop, aiming to combine deep reasoning with prompt responses and natural turn-taking.
Two components underpin its perception and interaction. Deep Perception is intended to capture rich acoustic cues in order to interpret user intent, while Seamless Duplex models synchronized audio streams so the system can handle pauses, backchannels, and interruptions naturally.
The report highlights Think-While-Speaking as the mechanism that resolves the tension between deep deliberation and latency, by executing private reasoning in parallel with spoken delivery. In reasoning mode, StepAudio 3 is reported to reach a 73.0 macro average on StepAudioChat. With Think-While-Speaking enabled, it is said to achieve dialogue and reasoning performance comparable to dedicated reasoning models while speaking in real time.
An integrated Voice Agent is also described as handling asynchronous tool execution without disrupting the dialogue flow. The report cites top-tier results across several dimensions: 90.6 on the MMSU benchmark, 98.9 Overall on the Artificial Analysis Full-Duplex Bench, and a 56.0% macro task-success rate on τ-Voice.