Can a single model learn when to speak and respond?

Cascaded conversational systems chain separate modules — VAD, ASR, language, TTS, animation, video generation — and pay for it twice: pipeline latency and error accumulation across handoffs. Wan-Streamer argues the cascade also makes the interactional problem unsolvable in principle, because deciding when and whether to respond, manage a turn, or absorb an interruption is not any one module's job. By representing language, audio, and video as one interleaved stream of input and output tokens under block-causal attention, it makes perception, generation, response timing, turn management, and cross-modal synchronization all learned jointly within a single Transformer, with streaming units as short as 160 ms at 25 fps and sub-second latency.

The load-bearing reframe: in a cascade, turn-taking has to be engineered (a VAD threshold, a silence timer); in a unified causal stream, timing is emergent behavior the model learns from the same sequence it uses to perceive and speak. Human interaction is full-duplex — we watch, listen, speak, and interrupt with overlap — and that overlap is precisely what module boundaries destroy.

This extends a thread the vault holds on collapsing the speech cascade. Since Can skipping transcription make voice assistants faster?, removing the ASR/TTS hops already buys dramatic latency; Wan-Streamer pushes the same logic one tier further, absorbing video and — crucially — turn control into the unified model rather than only the speech path. It also concretizes a long-running design question on the architecture map: since Why do AI conversations reliably break down after multiple turns? asks why multi-turn interaction degrades, full-duplex streaming reframes the turn itself as a learned, continuous decision rather than a discrete handoff between modules.

The strongest counterargument is the classic monolith-versus-pipeline trade: a single end-to-end model sacrifices the modular debuggability, independent upgrade paths, and component-level guarantees that cascades give you, and v0.1 is an early system whose latency/quality numbers are demonstration-scale, not field-hardened. The unification may simply move error accumulation inside the model where it is harder to inspect.