Two ways to talk back
Your scene partner can run on either engine. Both hear your actual voice. The difference is what happens next: one model answers out loud, or one model writes the line and a second one reads it.
Speech-to-speech
GPT Audio
- 1.GPT Audio hears your voice and answers in its own
- Starts talking
- 1.7 s
- after you stop (median)
- Cost
- 1.00¢
- per partner line
- First audio at the server
- 0.7 s
- At the server, slowest 10%
- 0.9 s
- Heard by players, slowest 10%
- 2.3 s
- Partner’s score from the judge
- 4.5 / 5 (6 scenes)
- Played so far
- 21 scenes, 60 lines
Speech-to-text-to-speech
Gemini + ElevenLabs
- 1.Gemini Flash-Lite hears your voice and writes a line
- 2.ElevenLabs reads it aloud
- Starts talking
- –
- after you stop (median)
- Cost
- 0.45¢
- per partner line
- First audio at the server
- 1.4 s
- At the server, slowest 10%
- 1.5 s
- Heard by players, slowest 10%
- –
- Partner’s score from the judge
- –
- Played so far
- 1 scene, 2 lines
How these numbers work
“Starts talking” is measured in players’ browsers, from the end of their line to the first sound of the reply. It includes the short pause that tells the game you’ve finished. “First audio at the server” leaves out the network and that pause. Both are medians over every partner line played here, test runs included.
Costs are per partner line, measured on typical lines. The judge’s partner score comes from the same Toastmasters-style evaluation players get, so it reflects how well each engine plays the scene, not how it sounds.