Automatic speech recognition for kreol morisien: A case study for the health domain

Abstract

Automatic Speech Recognition (ASR) has revolutionized human-machine interactions as it allows the use of speech as an input modality. Speech is easy, natural and it is a skill that most people possess in their respective languages. Therefore, speech technology contributes to the usability and inclusivity of applications. ASR in languages such as English is extensively developed as there are large amounts of relevant resources available such as audio or transcribed data. For languages which are under-resourced, such as Kreol Morisien, ASR is a monumental task. In this paper, an attempt at developing an ASR system in Kreol Morisien is described. The ASR system was developed for the health domain to enable the automatic recognition of medical symptoms in spoken Kreol. The data collection process included the manual creation of a list of 848 symptoms along with 4000 audio files. Using the created corpus, the acoustic model for Kreol recognition was built and trained. This paper also describes a user evaluation which was conducted in different environments. Findings showed that the accuracy of the acoustic model was mainly affected by the level of noise. The gender of the speaker and the pronunciation style (depending on the region where the speaker originates from) did not cause any significant difference in the performance of the acoustic model.