Speech-duplex conversations voiced by the new inference engine
Listen to each conversation and read along. The transcript is the engine's own word timing for what each side said, compared with the script it was given: it does not use Whisper, so you can judge by ear whether the voice says the script.
How to read a card
Audio: left ear speaker A = the person (user side, channel 0), right ear speaker B = Hera (assistant side, channel 1), cut at script end + 1.0 s.
Transcript vs script: green = the model said it but it is not in the script; struck = in the script, not said. Click a line to play from there (on a pair, both engines jump to the same time). Keys: left / right arrows skip 5 s.
Gate verdict (top of each player): pass, or the failures in red. cov = share of the worse side's script words said in order; tone_s = seconds of a stuck tone; listening_max = loudest sound while the side should be only listening; skipped_lines is added when cov is under 0.70.
Input (fold-out): the persona / system prompt and script lines each side was given.