Echo & Audio
If the bot hears itself, choose "Native". If audio is too quiet, try "Force loudspeaker".
Realtime VAD (only affects realtime model)
Lower = more sensitive (picks up quiet speech). Higher = bot ignores echo / background noise. Default 0.50.
How long you must be silent before bot responds. Higher = less false interruption, more latency. Default 500.
Audio kept before VAD detects speech. Higher = catches clipped onsets ("ej" instead of "hej"). Default 300.