Connected-digits recognition for an under-resourced language using Hidden Markov Models

This paper presents the development of a speech recognition system for automatically recognizing fluently spoken digit strings in Northern Sotho. The digit strings can be isolated or connected/continuous with known or unknown length. The digit recognition system has been trained with the aim of satisfying its potential end-users. Our main research focus was to enhance the robustness of a connected-digits recognizer such that it can handle continuous speech input restricted to numeric digits vocabularies. The Hidden Markov Model Toolkit (HTK) was used for experimentation. The standard technique that is based on the use of hidden Markov models (HMMs) was augmented with Cepstral Mean Vector Normalization (CMVN); a technique designed to handle convoluted distortions with the aim of increasing the robustness of speech recognition systems. A 1255 words dataset extracted from an existing general-purpose Northern Sotho speech database collected from mother tongue speakers between the ages of 16 and 60 was used in our experiment. The CMVN technique obtained a phone recognition accuracy of 75.84% and a word recognition accuracy of 62.30% whereas the standard HMM-based technique obtained phone recognition accuracy of 72.45% and a word recognition accuracy of 4.57%.