Performance of Speaker-independent Speech Recognisers for Automatic Recognition of Australian English

This paper investigates the performance of three speaker-independent speech recognisers (SISRs) that support continuous speech and are currently available for speaker-independent recognition of English. These speech recognisers were tested using a subset of the Australian National Database Of Spoken Language (ANDOSL) for the recognition of digits, sentences and a short paragraph under the “clean” data condition. The recognition accuracy was evaluated by two sets of tests, one set using a limited number of speakers (i.e. 13 speakers) and the other set using 75 of the speakers in the ANDOSL database. This paper describes the preparation taken to compare the recognisers operating under similar conditions. The paper does not reveal the name of the speech recognisers because the study reported here was limited. Nonetheless, the aim of the evaluation study was to ascertain which of the recognisers would be most suitable for applications in a meeting room where a sub-set of the vocabulary is selected as the active vocabulary for control of devices and for issuing commands. The variables investigated in this study were form of speech, vocabulary size, age and gender of the speaker. The contribution of this paper to speech science and technology is in its comparative evaluation of three speaker-independent speech recognisers., and in its use of a database containing utterances with Australian English accent.