High-precision medical speech recognition through synthetic data and semantic correction: UNITED-MEDASR