The automatic recognition of Afrikaans stop consonants in continuous speech by machine

The stops are modelled at a subphoneme level using continuous hidden Markov models. Each state within a particular Markov model represents a specific segment that might occur within a stop, for example, silence. These models, as well as being able to identify an unknown stop, can provide a `fine transcription´ used to obtain further features pertinent to stop recognition, such as voice onset time, to modify the initial classification provided by the stop model. This is achieved with a k-nearest neighbour probability density function estimator. The use of the system in recognition experiments gave results of 58% on stops found in most environments and 72% on stops found in the specific environment vowel-stop-vowel