An Investigation Into Discriminative Training of Input Transformations for Continous Speech Recognition

INFRODUCI'IONThe most successful continuous speech recognition systems use statistical techniques such as hidden Markov models (HMMS) to model and thereby recognise speech.The success ofthese techniques is based upon the amilabilitt-of powerful and tractable techniques for automatic parameter reestimation and speech recognition.Most of these speech recognition systems operate in the space of log power spectra, although the data may have undergone some simple, mostly linear.transformations such as frequency scale transformations (Mel-scale warping), cosine trans-