Voice Paradigms

Three paradigms of voice architecture: (1) cascade – a pipeline of VAD→ASR→LLM→TTS, modularity, but accumulation of latency and information loss at the joints; (2) end-to-end multimodal model (Omni) – a single model listens-thinks-speaks, lower latency, prosody is preserved, but still “turn-based” by VAD; (3) full duplex – the model listens and speaks simultaneously, deciding dozens of times per second whether to speak/listen/be silent/interrupt, without assuming turn-taking. A common thread through all three: how to get rid of guessing turn boundaries by silence.

Related: [Cascaded Voice Pipeline], [Full Duplex], [Streaming Speech Perception]