Rapid adaptation using linear spectral transformation for embedded speech recognisers

Embedded speech recognisers are typically used in unknown mobile environments where the acoustic conditions frequently change. Since a large amount of adaptation data is not usually available for such environments, the adaptation methods for the acoustic models of these recognisers must improve the recognition performance with only a small amount of adaptation data. In this Letter, we show that maximum likelihood linear spectral transformation provides the advantage of rapid adaptation using a very limited amount of adaptation data for the embedded acoustic models.