Developing an AI-Powered Tool for Dialect Identification in Multilingual Speech Transcripts

Thesis Abstract

Effective communication in multilingual contexts relies heavily on accurately identifying regional dialects within speech transcripts, yet current methodologies often lack automation and scalability, limiting their utility in large-scale linguistic analysis and language technology applications. This study aims to develop an AI-powered tool capable of automatically identifying dialectal variations within multilingual speech transcripts, thereby enhancing linguistic resource development, improving speech recognition systems, and advancing sociolinguistic research. The primary objectives are to (1) examine the linguistic features that distinguish dialects in speech transcripts, (2) design and implement a machine learning model integrating phonetic, lexical, and syntactic cues, and (3) evaluate the tool's accuracy and robustness across diverse multilingual dialect datasets. The research adopts a mixed-methods design, combining quantitative and qualitative approaches. The population of the study comprises audio speech recordings and corresponding transcripts collected from a corpus of 1,200 multilingual speakers across three distinct regions, representing three different languages and dialectal varieties. A stratified random sampling technique will be employed to select 300 speech samples per region, ensuring representative coverage of dialectal diversity. Data collection will utilize existing speech corpora augmented with field recordings, where necessary, using high-fidelity microphones, and transcripts verified through linguistic annotation. Methodologically, the study involves feature extraction of phonetic, lexical, and syntactic markers through automated speech processing tools. Machine learning classifiers—specifically convolutional neural networks (CNNs) and support vector machines (SVMs)—will be trained and validated on a segmented dataset, with cross-validation techniques employed to mitigate overfitting. The model's performance will be assessed using standard metrics such as accuracy, precision, recall, and F1 score, and its interpretability evaluated through feature importance analysis. The study will incorporate the theoretical framework of Speech Community Theory and Markedness Theory to inform feature selection and model development, positing that dialectal variation can be systematically represented through identifiable linguistic features rooted in sociolinguistic constructs. Expected findings include high classification accuracy (above 85%) in dialect identification across the studied languages, with particular emphasis on phonetic markers as primary discriminators. The study anticipates demonstrating that machine learning models can reliably distinguish dialectal features from speech transcripts, even in multilingual and code-switching contexts. Additionally, the research expects to reveal contextual variations in dialectal markers linked to socio-cultural factors, validating the theoretical underpinnings of sociolinguistic variation. The contribution to existing knowledge lies in pioneering an integrated AI-based approach to dialect identification within speech transcripts in multilingual settings, addressing gaps in scalable, automated linguistic diagnostics and enhancing tools for sociolinguistic, forensic, and language technology applications. The study advances the understanding of linguistic feature representation in machine learning models and sets a foundation for future research on dialectal variation across lesser-studied languages. In conclusion, the developed tool will provide a reliable, efficient, and scalable means of dialect identification, with potential applications extending into automated transcription services, language preservation, and sociopolitical studies. Recommendations include further refinement for low-resource language contexts, integration with speech recognition systems, and expansion to include socio-pragmatic dialect features. Future research should explore longitudinal data to examine dialectal change over time and the integration of multimodal data to improve model robustness. This study affirms the viability of combining sociolinguistic theory with advanced machine learning techniques to address critical gaps in multilingual dialect recognition, fostering progress in linguistic resource development and language technology innovations.