Speech-to-speech. S2S, models collapse the voice pipeline into a single end-to-end neural network. Audio goes in, audio comes out. In theory the latency drops dramatically and the prosody control becomes seamless because the model never has to round-trip through text.
In 2026, the current generation of S2S models is impressive in demos and disappointing in production. Tool-calling reliability is worse than the pipeline approach. Handoff quality is worse. Brand-tone consistency is worse. The latency advantage is real but the quality regression is real too.
Most serious voice-AI vendors, including us, believe S2S becomes the right answer in 12-18 months as the tool-calling layer matures. Until then, the STT-LLM-TTS pipeline with careful engineering is the production-grade architecture. Vendors who shipped S2S today usually had to give something up to do it.