I usually listen to a few voices on the provider site first (like ElevenLabs), then bring the best ones into VAPI and try them in real call flows, because that’s when you really notice what works and what doesn’t