SRA System Design: Using Deep Learning to Analyse Experimental data for Speech Researchers

Abstract

Speech researchers expend tremendous effort when measuring and analyzing audio data from experimental participants and evaluating large amounts of audio data to aid speech research. In traditional assessment methods, researchers listen verbatim to long audio files from participants to find the keywords needed for the assessment. This approach is very tedious and time intensive for speech researchers. To improve research efficiency and efficacy, we designed a system called SRA (Speech Recognition Assistant) based on the DeepSpeech model. SRA effectively supports speech researchers in the evaluation of participant experimental audio data. The purpose is to achieve the following: (i) simplify the workflow for analyzing data and (ii) reduce the time cost for researchers to extract data; (iii) provide a framework with potential for adaptation for use in other fields, such as speech science, audiology, and hearing science.