Developing a Neural Network Model for Real-Time Dialect Identification in Speech Transcripts

Thesis Abstract

The rapid proliferation of digital communication tools has underscored the necessity for advanced natural language processing (NLP) systems capable of accurately identifying dialectal variations in speech transcripts in real time. Dialectal diversity presents significant challenges for speech recognition and translation systems, often leading to misinterpretations and reduced accessibility for diverse language communities. This study aims to develop a robust neural network (NN) model for real-time dialect identification, enhancing the fidelity of speech-to-text applications across dialectal variants. The specific objectives include (1) to analyze existing dialect detection methodologies; (2) to design and train an optimized neural network architecture tailored for dialect classification; (3) to evaluate model performance through empirical testing; and (4) to assess the model’s practical applicability in various speech recognition contexts. The research adopts a quantitative, experimental design grounded in machine learning principles, specifically utilizing supervised learning techniques. The population comprises annotated speech transcript datasets representing five major dialect groups within a linguistically diverse region. A stratified random sampling approach facilitated the selection of a balanced sample of 2,000 speech transcripts, with 1,600 allocated for training and 400 reserved for testing. Data collection involved sourcing speech transcripts from linguistic repositories, digital speech corpora, and annotated datasets provided by regional language authorities. A custom-developed dataset was curated to ensure balanced representation of dialects, phonetic variations, and speaker demographics, thereby enhancing the model’s generalization capabilities. The neural network architecture was designed using multiple layers of convolutional and recurrent units, leveraging bidirectional Long Short-Term Memory (BiLSTM) layers to capture temporal dependencies in speech features. Preprocessing involved extracting Mel-Frequency Cepstral Coefficients (MFCCs) and spectral features from speech transcripts, which served as input vectors for the model. Model training employed the Adam optimization algorithm, with categorical cross-entropy as the loss function, over 50 epochs, and validation was performed using k-fold cross-validation to mitigate overfitting. To evaluate the model’s accuracy, precision, recall, and F1-score metrics were computed; additionally, Receiver Operating Characteristic (ROC) and Area Under Curve (AUC) analyses provided insights into classification performance at different thresholds. Expected findings include high classification accuracy (anticipated at over 85%), with detailed insights into the importance of specific acoustic features for dialect distinction. The neural network model is projected to outperform existing machine learning approaches such as Support Vector Machines (SVM) and traditional Hidden Markov Models (HMM) in terms of speed and accuracy, offering real-time classification capabilities suitable for integration into speech recognition systems. The study also anticipates revealing the influence of phonetic and prosodic features on dialect identification, contributing to a more nuanced understanding of dialectal variations in speech processing. This research significantly contributes to the body of knowledge by demonstrating the viability of deep learning-based models for real-time dialect detection, providing a scalable and adaptable framework that can be extended to other linguistic communities and languages. The theoretical foundation draws on the Cognitive Load Theory and the Speech Paralinguistic Framework, which inform the understanding of how acoustic cues contribute to dialectal differentiation. Practical implications include improved accuracy in dialect-specific speech recognition, enhancing user experience in multilingual digital environments, and informing policy development for linguistic preservation and technological inclusion. The study concludes that neural network models hold promise for bridging linguistic diversity with advanced speech processing applications. Recommendations advocate for further research into integrating multimodal data (such as visual cues) to enhance dialect recognition robustness, as well as the development of lightweight models suitable for deployment on mobile devices. Future studies should explore cross-dialect transfer learning and domain adaptation techniques to accommodate evolving linguistic landscapes and dialectal shifts, ultimately fostering more inclusive and accurate speech-driven technologies.