A novel speech reconstruction algorithm for DSR back-end