A N-gram approach to overcome time and parameter independence assumptions of HMM for speech recognition

There is significant interest in developing new acoustic models for speech recognition that overcome traditional HMM restrictions. In this work, we propose to use a N-gram based augmented HMM. Two approaches are presented. The first one consists on overcoming the parameter independence assumption. This is achieved by modelling the dependence between the different acoustic parameters, using N-gram modelling. Then, the input signal is mapped to the new probability space. The second proposal tries to overcome the time independence assumption, by modelling temporal dependencies of each acoustic parameter. Different configurations have been tested, results showing that adding long span information is beneficial for ASR performance.