In-ear-voice: towards ultra-low power Audio ML