Choosing TTS for a real time AI call is not a beauty contest
A builder note on why voice quality, latency, controllability and failure behaviour must be judged together.
The best TTS voice in a demo is not necessarily the best TTS system in a live conversation; production choice is a systems decision.
When people compare text to speech systems, they usually start by listening to sample voices. That is useful, but it is not how a production voice agent experiences TTS.
A live system cares about the whole loop.
Start with the conversation, not the voice
The user finishes speaking. Speech recognition has to decide that the turn is over. The model has to respond. TTS has to generate useful audio. The first sound has to reach the caller before the silence becomes awkward.
A voice that sounds exceptional after three seconds can lose to a slightly less expressive voice that starts naturally and predictably.
What I would measure
I care about four things together: first audio latency, streaming stability, controllability of pace and emphasis, and what happens when the user interrupts.
Cost matters too, but only after the system clears the experience threshold.
The broader builder lesson
AI product choices are increasingly systems choices. A component can be best on its own and still be wrong for the product because the product is experienced end to end.