Differential Approach to Acoustic Model Adaptation

This paper discusses a ‘differential approach’ to acoustic model adaptation to different channel, noise and speaker conditions for speech recognition with a concept of vector field adaptation. Adaptation of acoustic model, such as HMM output probability densities, is modeled as a vector field in the acoustic feature vector space which effects on the models to move along the local vector directions. Specifically, if the feature vector is based on cepstrum, multiplicative channel characteristic is interpreted as a uniform vector field where the channel cepstrum vector is added to all feature distributions. Additive noise forms a vector field where 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. Experimentally compared with existing model composition approaches for noisy speech recognition, this approach drastically reduced the computational cost while providing equivalent recognition performance. Difference between speakers is also formulated as a vector field where only assumption is that the vector field is smooth from a physiological view. Vector Field Smoothing (VFS) technique for speaker adaptation is then discussed in this context.