Latest / Logosofia no Dia a Dia
How Speech-to-Speech Models Eliminate the Robot Voice
Why do so many AI voice agents still feel like talking to a robot? This episode unpacks the architectural difference between traditional pipeline systems (ASR → LLM → TTS) and the new class of natively integrated speech-to-speech models. We explore how the text bottleneck destroys prosody and emotion, why cumulative latency breaks conversational rhythm, and how models like OpenAI's Realtime API, Moshi, and Hume's EVI process audio end-to-end. We also cover the trade-offs between elegance and production readiness, and why pipeline tools still dominate despite their seams. Episode 2512 —…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.