Medium-vocabulary speech recognition for under-resourced languages

Abstract

We report on the development of speech-recognition systems that are able to perform accurate recognition on mediumvocabulary tasks (i.e. tasks that require distinctions between approximately 200 different terms). We are able to achieve error rates of less than 5% (our design goal) on four underresourced languages as well as English, by using training corpora that contain 70–100 hours of speech per language. The majority of the errors stem from words such as abbreviations, foreign words or names, which do not adhere to the standard orthography of the target language. We also find that recognition accuracy does not depend strongly on the number of occurrences of a term in the training set or the length of the term to be recognized, and that a few problematic speakers are responsible for a disproportionate number of errors.