On-line speaker adaptation on telephony speech data with adaptively trained acoustic models