Connected digit recognition in Afrikaans using hidden Markov models

Abstract

We implement and evaluate the performance of a state-of-the-art connected digit recognition system for spoken Afrikaans digits. The creation of a database for our purpose is discussed. Results indicate that an efficient implementation achieving high recognition accuracy ( 99.2% digit accuracy and 93.3% string accuracy) is attainable. Results also indicate that improvement due to duration modelling is confined to cases where there are relatively few states per hidden Markov model (HMM). The HMM is used to model the statistical properties of speech signals, as it is well understood and forms a main com­ponent of almost all state-of-the-art automatic speech recognition systems. Continuous models, in particular multi-modal Gaussians, are used to model speech feature distributions. Experiments are performed on a database consisting of' spoken telephone numbers in Afrikaans from 33 speakers. An HMM is initialised for each digit from labelled isolated digits and embedded re-estimation is used to further train models on continuously spoken telephone numbers. Parametric models, in the form of the Gamma distribution, are used to model word and state duration density. Duration modelling is incorporated as part of the recognition process. Classification experiments are performed on continuous digit test data and performance is measured for a range of modelling parameters.