Direct Speech-to-Speech Agents

StepAudio 3 combines a continuous listen-converse-think-act loop, parallel reasoning and asynchronous tools; Gemini Live independently emphasizes parallel reasoning during live voice tasks. This strengthens the existing direct-audio architecture line, but does not prove that every implementation removes transcription. Independent latency comparisons and repeated production workflows remain necessary; the line is still below the five-publication threshold.

An open voice-agent stack now sends audio directly to a multimodal language model instead of converting speech to text first. Removing transcription could preserve prosody, simplify latency-sensitive pipelines and change the economics of real-time voice products. This is one implementation, so the line remains on watch until independent stacks demonstrate lower latency, better task performance or production adoption without a separate STT stage.