Use of vector-valued dynamic weighting coefficients for speech recognition: maximum likelihood approach