Performance improvement of CSR using a segmental-feature HMM

Abstract

Despite several decades of research activity, speech recognition still retains its appeal as an exciting and growing field of scientific inquiry. The goal of automatic speech recognition is to develop techniques and systems that enable computers to accept speech input. To accomplish the speech recognition, the input speech signal, via a microphone or telephone, is first transformed into a set of useful measurements or features at a fixed rate. These measurements are used to create a pattern representative of the features, or to generate templates or models for the reference patterns in training step. In the recognition step, these features are also used to find the most likely word candidate. If the reference patterns are characterized by the statistics of the features, training data are used to determine the model parameters. In a statistical framework, an acoustic model means an inventory of elementary probabilistic models of basic linguistic units to build word representations. Therefore, the feature measurements and the acoustic models have an important role in speech recognition system. A Hidden Markov model (HMM) is a representative of an acoustic modeling and is the predominant and current best performance speech recognition algorithm. Even though an HMM shows good results in modeling the statistical variations of acoustic speech signals, it is reported that some of its assumptions are not appropriate in practice. Thus, various studies are presented to relax some weakness of HMMs in the feature representations and acoustic modelings. In this point of view, we presented a new feature measurement to represent the set of frame features in detail and an acoustic model for characterizing the proposed features, and developed an algorithm based upon a general framework of HMMs. The proposed feature measurement uses a set of frame features rather than single frame feature because single frame feature cannot describe the temporal dynamics of speech signals. A s...