It makes use of phoneme lattices generated by a phoneme recognizer.Acoustic Model (AM) based CM is also introduced.For a decoded speech frame aligned to an HMM state, the CM based on AM is calculated.