Automatic annotation of reading using Speech Recognition: A pilot study

Abstract

Speech is a time-varying continuous stimulus. Analyzing a speech sample is generally done by transcribing it which can be considered to be a time-consuming process. A time annotated transcription is required for a detailed evaluation of the speech sample to study the language and prosodic features. The focus of this paper is to develop a tool to automatically segment audio recordings into silences and chunks of speech corpus and further evaluate the same using speech recognition technology and therefore, obtaining a time annotated speech transcript. Speech samples were obtained from a group of 20 normal adults whose age ranged between 18 and 22 years. The speech sample contained bilabial phrases of 2–3 words in length. An endpoint-based silence removal was designed, and the extracted speech cluster was further analyzed for features, and HMM models were used for cluster modeling during the training and testing phase. The tool was developed using MATLAB platform, and PRAAT software was used to perform ground truth testing. The experimental results revealed a high recognition rate of around 90% with a time annotation difference of 0.5 s from that of ground truth testing (PRAAT analysis) and the developed tool transcript (MATLAB). The tool was able to successfully segment speech and transcript the same with high accuracy. Further, the tool can be refined to analyze and transcribe spontaneous speech and in the future can be applied to the field of communication sciences for more accurate diagnostic and management procedure.

Keywords: Annotation, Speech, recognition, communication, sciences

Cite this Article

Akshay Mohan Mendhakar, Sangeetha Mahesh, Automatic annotation of reading using Speech Recognition: A pilot study. Research & Reviews: A Journal of Bioinformatics. 2018; 5(2): 25–29p.