This paper proposes a speaker adaptation algorithm that covers a wide range of adaptation data. The parameter smoothing technique improves adaptation performance for a small amount of adaptation data; however, this smoothing usually reduces adaptation efficiency for a large amount of adaptation data. Our method dynamically controls smoothing strength by using information on the amount of adaptation data to achieve good adaptation performance over a wide range of adaptation data. The proposed method is combined with the maximum a posteriori (MAP) estimation technique, and its effectiveness is shown on a Japanese 26 phoneme recognition test.
Bibliographic reference. Tonomura, Masahiro / Kosaka, Tetsuo / Matsunaga, Shoichi / Monden, Akito (1995): "Speaker adaptation fitting training data size and contents", In EUROSPEECH-1995, 1147-1150.