Design, implementation, and evaluation of a spoken language identification system for multilingual university settings

Thesis Abstract

In multilingual university settings, spoken language identification (LID) systems are essential for real-time classroom assistance, routing of student inquiries, and secure access control, yet existing solutions underperform when confronted with culturally diverse accents, code-switching, and domain-specific vocabulary. This study aims to design, implement, and evaluate a robust LID system tailored to the linguistic diversity of university environments, with the objective of achieving high accuracy across languages commonly encountered in higher-education contexts and providing reliable confidence estimates for downstream applications. Specifically, the objectives are (1) to develop a modular LID architecture combining acoustic-phonetic and end-to-end neural components; (2) to create and curate a university-domain speech corpus comprising five target languages (English, Mandarin, Spanish, French, and Arabic) with balanced intra-language variability (gender, age, native/non-native speakers) and realistic conversational and lecture-style recordings; (3) to implement language normalization and robust handling of code-switching and background noise; (4) to evaluate system performance against state-of-the-art baselines using comprehensive metrics and statistical tests; and (5) to analyze system interpretability and deployment readiness for university IT ecosystems. The research adopts a mixed-methods, design-implement-evaluate paradigm, drawing on the theoretical framework of probabilistic modeling for LID and the Information-Theoretic Perspective on language cues, supplemented by the Theory of Sociolinguistic Variation to account for intra-language diversity. A corpus of 200 hours of speech will be collected, including 40 hours per target language, with 80 hours from native speakers and 60 hours from proficient non-native speakers, and 20 hours of naturalistic classroom recordings. The dataset will be split into training (70%), validation (15%), and testing (15%) sets, ensuring speaker-disjoint partitions. Data collection instruments include standardized speech elicitation protocols, synchronized metadata schemas, and noise augmentation tools to simulate cafeteria, outdoors, and hallway environments. The methodology integrates a hybrid LID pipeline an acoustic-phonetic frontend grounded in i-vector/Probabilistic Linear Discriminant Analysis (PLDA) representations, and a neural backend leveraging a Transformer-based encoder with self-supervised pretraining (wav2vec 2.0) fine-tuned for language identification. A language-agnostic spectro-temporal feature extractor will be employed to capture cross-language cues, complemented by phonotactic modeling to improve discrimination in closely related languages. System calibration will use temperature scaling to produce reliable confidence scores. The primary evaluation metrics will include language identification accuracy, equal error rate (EER), log-likelihood ratio (LLR) calibration, and macro-averaged F1 scores across languages, with secondary analyses of per-language performance, robustness to code-switching, and resilience to background noise at signal-to-noise ratios (SNRs) of 0–20 dB. Statistical analyses will employ repeated-measures ANOVA to compare pipeline variants, and pairwise McNemar tests for significance on the testing set. Interpretability will be assessed via saliency mapping and analysis of attention weights to identify language-specific cues. Expected findings include (i) the hybrid LID architecture achieving superior accuracy relative to baselines (target overall accuracy > 92%, EER < 6% across languages) and improved resilience to code-switching; (ii) demonstration that incorporating phonotactic features and language-specific phoneme inventories significantly boosts discrimination for closely related languages (e.g., English-Spanish or Arabic-French); (iii) evidence that confidence estimates align with true correctness, enabling threshold-based routing for classroom support tools; and (iv) practical guidance on data collection protocols and deployment considerations for university IT infrastructures. The study contributes to knowledge by providing a rigorously validated, scalable LID framework that accounts for multilingual university dynamics, offers reproducible corpus-building methodologies, and informs best practices for deploying language-aware educational technologies. The main conclusion anticipates that a carefully engineered hybrid model, with explicit phonotactic components and calibrated confidence outputs, can deliver reliable language identification in real-world university settings, enabling improved accessibility, efficient resource routing, and enhanced security. Recommendations include expanding the language set to capture regional dialects, integrating user-adaptive models that personalize to institutional cohorts, and developing open benchmarks for university-domain LID to foster further research and cross-institutional collaboration.