Bohan Li presents a cascaded voice-agent architecture with separate transcription, language-model planning and speech-generation stages. He combines fast streaming transcription with slower corrective transcription, allowing updated recognition to replace stale context rather than treating every early transcript as final.
Bohan Li moves tool work into background execution and injects completed results into the conversation. Early response generation can be cancelled and restarted when relevant information changes. The harness waits for the user's turn to end before emitting speech, balancing computation overlap against the risk of interrupting the user.
Bohan Li describes caching spoken prefixes while generating the personalized remainder of a response. Speech generation receives the whole response for prosody, with duplicate introductory audio suppressed. He acknowledges possible joins or small audio hiccups rather than claiming that cached speech is always seamless.
Bohan Li illustrates the approach with a prerecorded clinic-booking conversation involving pauses, an unavailable appointment and a suggested alternative. The demonstration is an example of conversational behavior, not an independent latency benchmark or proof of a live clinical backend. Closing recruitment and company promotion are omitted.
Watch on YouTube
