Evidence types
- Published specification: a capability or limit described by a provider or project.
- Vendor measurement: a reported experiment with its model, clock, and conditions retained.
- Site measurement: a test performed by this site, published only with its protocol and evidence.
Current coverage
The site compares nine hosted APIs and three self-hosted or animation tools. A retired Hedra realtime service remains in historical notes on the homepage. Current charts exclude retired services.
Pricing rules
Rates retain their currency, mode, model, and subscription basis. Unknown rates remain unknown. Entry usage quotes and calculated monthly totals are presented separately.
Latency clocks
- Voice-to-voice: end of user speech to first reply audio.
- Complete response: use the exact event definition reported by the source.
- Rendering: audio or another model input to generated video.
- First frame or first byte: preserve the reported input event and delivery boundary.
- FPS: processing or stream throughput, measured separately from latency.
Editorial order
The default order considers integration fit, creative control, and documented capabilities. A position is an editorial recommendation. Measured quality and speed require comparable test evidence.
Source reviews
Sources were last reviewed Oct 13, 2026. The four-day review schedule checks published information before advancing that date. Original publication dates remain fixed; substantive changes have a separate update date.
Benchmark status
This site has not completed an independent cross-provider experiment. Its voice chart reproduces approximate readings from LemonSlice's April 2026 report; the SyncNet comparison was run by D-ID. Historical versions and vendor limitations stay attached to these results.
A controlled evaluation
- Use consented avatar inputs, identical audio, and named model versions.
- Declare timing events, region, network conditions, output settings, and session mode.
- Separate cold starts, warm sessions, interruptions, and sustained usage.
- Use 30 turns as a pilot, then increase samples before strong percentile claims.
- Preserve failed sessions and publish recordings, run metadata, and uncertainty.