- Smart Turn now takes into account `vad_start_seconds` when buffering audio,
  meaning that the start of the turn audio is not cut off. This improves
  accuracy for short utterances.

- The default value of `pre_speech_ms` is now set to 500ms for Smart Turn.