45-64
A general description of automatic speech recognition systems architecture
Authors: Valentina Sofroni, Alexandru Stan
Number of views: 399
Over the last decades, the progress in the ASR domain has been amplified by a significant amount of technical and scientific advancements, amongst which the continuous expansion in the power of computing systems. From a technological point of view, speech recognition has been undergoing tides of major innovations in methodology, algorithms, learning concepts or practical system implementations. This paper provides an up-to-date perspective on the architecture of automatic speech recognition systems and their constituent components. It presents modeling paradigms currently dominant in this type of systems (Hidden Markov Models, Gaussian mixture models, Bayes classifiers, N-gram language model, etc.) together with the architectural constraints they impose upon the design of the system. This study stands for an intermediate step in a larger process which aims to conceive and implement a highly accurate speaker independent ASR system for the recognition of the Romanian language in a limited field of application, such as justice.