What the current charts establish
Published latency figures can help identify what to test, but different clock definitions prevent a controlled fastest-provider result. The homepage keeps full responses and video rendering in separate views.
Historical voice-to-voice comparison
LemonSlice's April 15, 2026 report uses 50 turns per provider with a common reference speech stack. The chart uses approximate figure readings rounded to 0.1 seconds. Phoenix-4 and CARA-3 are historical versions.
Lip-sync evidence
D-ID publishes a SyncNet comparison using five avatars per provider. Competitor snapshots and their equivalence to current models remain unspecified here. Anima was not included.
What to measure in your trial
- End of user speech to the first audible reply.
- Incoming audio to its corresponding video output.
- Interruptions and recovery, with failed runs retained.
- Delivered FPS and output dimensions over sustained sessions.
- Actual billed minutes, starts, subscription allocation, and cost.
Public demo checks
A separate walkthrough records access controls, model labels, and vendor sample playback. It leaves measured performance fields blank.