Jacobian adaptation of noisy speech models
Shigeki Sagayama, Yoshikazu Yamaguchi, Satoshi Takahashi · 2002
A Jacobian approach to fast adaptation of acoustic models is described. Acoustic models of speech under assumed noise and channel condition A are compensated by Jacobian matrices with the difference between condition A and actual condition B. Compared with existing model composition approaches for noisy speech recognition, this approach drastically reduces the computational cost while providing equivalent recognition performance. Extension of this analytic approach to acoustic model adaptation is also extensively discussed.