Episode Details

Ever notice how AI-generated podcast dialogue feels slightly off—too clean, too polite, with zero interruptions? This episode pulls back the curtain on our own AI voices and the "realism gap" in text-to-speech. We explore why current TTS models are trained on single-speaker studio audio, why human conversation is full of overlaps, backchannels, and two-hundred-millisecond gaps, and how multimodal models might finally teach machines to be convincingly messy.