Acoustic Model Adaptation with Multiple Supervisions

This paper reports on several experiments performed during the development of the ITC-irst transcription system for the TC-STAR '06 evaluation campaign. The aim is to nd methods of exploting a set of alternative hypotheses produced by different systems to derive a transcription that is more acccurate than any of these. The most used technique for combining alternative hypotheses relies on the ROVER technique. In this work we found that another approach, based on adaptation of a reference system, may provide some advantage over that technique.