Towards an Optimal Phonetic Representation of the Acoustic Signal as Starting Point for Search in Hsr Models

The research presented in this paper is concerned with the development of an automatic phone recogniser (APR) that converts an acoustic speech signal into a phone string, that meets the criterion that the number of phones should be ‘correct’. Two methods are investigated, viz. the optimisation of the APR on the ‘number of phones’ criterion and the introduction of variants in the phonetic transcriptions of the words in the lexicon. Experiments on a city name corpus showed that the optimisation of the APR on the length criterion and the optimisation of the phonetic transcriptions both caused a decrease in phone deletions. However, only the optimisation of the phonetic transcriptions caused an overall increase of the performance of the APR. Finally, the effects of adding the variants to the lexicon of a standard ASR system are investigated.