The latter. You get billed on how many tokens the voice uses. The estimate is just a rough guide. If the assistant is just listening to the user for 1 minute, you wouldn't be billed for voice, as no voice tokens was used