A Voice-Based Personal Assistant for Mental Health in Kreol Morisien

Abstract

Voice-based smart personal assistants (VSPAs) are applications that recognize speech-based input and perform a task. In many domains, VSPA can play an important role as it mimics an interaction with another human. For low-resource languages, developing a VSPA can be challenging due to the lack of available audio datasets. In this work, a VSPA in Kreol Morisien (KM), the native language of Mauritius, is proposed to support users with mental health issues. Seven conversational flows were considered, and two speech recognition models were developed using CMUSphinx and DeepSpeech, respectively. A comparative user evaluation was conducted with 17 participants who were requested to speak 151 sentences of varying lengths in KM. It was observed that DeepSpeech was more accurate with a word error rate (WER) of 18% compared to CMUSphinx at 24%, that is, DeepSpeech fully recognized 76 sentences compared to CMUSphinx where only 57 sentences were fully recognized. However, DeepSpeech could not fully recognize any 7-word sentences, and thus, it was concluded that the contributions of DeepSpeech to automatic speech recognition in KM should be further explored. Nevertheless, this research is a stepping stone towards developing more VSPA to support various activities among the Mauritian population.