Speaker-Independent Automatic Speech Recognition System for Mobile Phone Applications in Punjabi

Abstract

Speaker-independent Automatic Speech Recognition (ASR) system based mobile phone applications are gaining popularity due to technological advancements and accessibility. Speech based applications may provide mobile phone accessibility and comfort to people performing activities where hand-free phone access is desirable e.g. drivers, athletes, machine operators etc. Similarly, users with disabilities like low vision, blindness and physically challenged may use it as an assistive technology. Development of ASR system for a specific language needs accurate, reliable and efficient acoustic model having language-specific pronunciation dictionary. Punjabi language is one of the popular languages worldwide having more than 150 million speakers. Three acoustic models- continuous, semi-continuous and phonetically-tied are developed based on three pronunciation dictionaries- word, sub-word and character based. Analysis of performance results validate Punjabi language principle “One word one sound” by having better accuracy and reliability for character based pronunciation dictionary than others. Further, phonetically-tied model outperforms others in terms of accuracy, word error rate and size due to reasonable number of Gaussians.