overhaul the way silences are being dealt with
now handled in each transcriber's realtime transcription method
every time a sound is heard the buffer is first populated with half a
second of the previous silence buffer, so that it is properly handled by
the AI models.

(this only apply to whisper and openaiapi backends)