So maybe perfect turn detection shouldn’t be the goal

July 15, 2025

> So maybe perfect turn detection shouldn’t be the goal.

I've had several conversations lately about a general version of this idea: don't define success for your AI system relative to a hypothetical "perfect" performance; instead benchmark against the strengths, weaknesses, and failure modes of human-to-human interaction.

This is true for turn detection. The goal isn't to never get an end-of-turn classification wrong. The goal is for the human to feel like the conversation flows naturally.

It's true for lots of other parts of voice AI interaction, too: pronunciation, instruction following, dealing gracefully with unexpected inputs, deciding when to automatically escalate to a human operator.

Rajiv Ayyangar@rajivayyangar

@kwindla @swyx Humans indeed do this all the time. And we get it wrong and either talk over each other or apologize and cede.

I saw some stat that in group conversation we talk over each other or laugh / make phatic sounds ~15% of the time. So maybe perfect turn detection shouldn’t be the