Boosting under-resourced speech recognizers by exploiting out of language data - Case study on Afrikaans

Under-resourced speech recognizers may benefit from data in languages other than the target language. In this paper, we boost the performance of an Afrikaans speech recognizer by using already available data from other languages. To successfully exploit available multilingual resources, we use posterior features, estimated by multilayer perceptrons that are trained on similar languages. For two different acoustic modeling techniques, Tandem and Kullback-Leibler divergence based HMMs, the proposed multilingual system yields more than 10% relative improvement compared to the corresponding monolingual systems only trained on Afrikaans.