Speech-to-Text Integration in Multilingual Mobile Apps

Abstract

Speech-to-text (STT) technology is increasingly embedded in mobile applications, yet reliable transcription remains difficult when users alternate languages, accents, scripts, and acoustic environments within the same interaction. This study investigates a mobile-oriented multilingual speech recognition framework designed to support dynamic language identification, code-switched speech, and computationally constrained deployment. The research gap lies in the limited integration of transcription accuracy, intra-utterance language switching, mobile inference latency, and resource efficiency within a unified evaluation framework. A language-aware speech-processing architecture is proposed in which multilingual acoustic representations are combined with adaptive language routing and contextual decoding. The intended experimental framework concentrates on five linguistically distinct languages—English, Hindi, Bengali, Tamil, and Telugu—with additional emphasis on English–Indic code-switching. Rather than evaluating recognition solely through conventional word error rate, the study considers language identification reliability, code-switch boundary performance, character-level accuracy, latency.