OPTIMISING SWAHILI DIGRAPH RECOGNITION USING A FEATURE VECTOR REDUCTION MODEL

Abstract

Recognising and classifying digraph letter pairs representing single phonemes remains a key linguistic and computational challenge in low-resource languages. Digraphs are essential for accurate pronunciation and comprehension in Natural Language Processing, speech and text recognition, transcription, and machine translation. Swahili, a widely spoken East African lingua franca, is a low-resource language with limited linguistic datasets and computational tools, featuring a rich phonetic and syllabic structure. Current models focus on high-resource languages and word-level processing, often neglecting the complexities of individual phonemes and digraphs. This study proposed an optimised model for Swahili digraph recognition by introducing a Feature Vector Reduction Model using Multinomial Logistic Regression, aimed at enhancing both computational efficiency and linguistic precision. The research pursued four key objectives: (1) phonetic and linguistic analysis of Swahili digraphs, (2) development of an annotated digraph corpus using a CNN-based model, (3) optimisation of digraph recognition using FRMLR, and (4) empirical validation of the FRMLR model. Anchored in positivist epistemology and employing the Design Science Research Methodology, the study integrates empirical validation, reproducibility, and measurable outcomes to advance NLP for low resource languages. A significant contribution of this study is the creation of the Annotated Swahili Digraph Corpus, comprising 31,197 lexicon entries annotated with digraph-vowel interactions and phonotactic rules. This multimodal corpus built using CNN-based extraction, incorporates both text and synthetic speech data and achieved a recognition accuracy of 92.5% (MAE = 0.112, R² = 0.89), thus filling a critical gap in Swahili linguistic resources. The FRMLR framework effectively reduced 39-dimensional Mel-Frequency Cepstral Coefficients into compact 5-dimensional probability vectors, resulting in 70% memory savings (from 1500 MB to 450 MB) and a 77.8% decrease in training times (from 45 minutes to 10 minutes). It consistently outperformed baseline methods such as Principal Component Analysis with Support Vector Machines, achieving a precision of 91.8% and a recall of 90.3%, with minimal misclassifications. Statistical validation confirmed the effectiveness of FRMLR. A one-way ANOVA test (p = 0.003) showed a statistically significant improvement in FRMLR’s performance compared to baseline models, Post hoc analysis using Tukey’s HSD test further validated that FRMLR significantly outperformed Logistic Regression (p = 0.002), SVM (p = 0.015), and KNN (p = 0.001), while performing comparably to CNN (p = 0.670). The AUC-ROC analysis consistently produced values exceeding 0.91 across the training, validation, and test datasets. This reinforces the model’s strong ability to distinguish between phonetic patterns. The low false positive rates (FPR: 0.07–0.10) further confirm FRMLR's precision in minimising misclassifications. This quality is crucial for applications that require high-confidence predictions, such as educational tools and speech recognition systems. This study enhances computational linguistics by integrating machine learning with natural language processing for ASR in low-resource contexts. The FRMLR model effectively reduces phonetic feature dimensions, and CNN aids in corpus construction and pattern recognition. These approaches establish scalable benchmarks for transcription accuracy and voice systems, promoting digital inclusivity while striking a balance between linguistic rigour and computational efficiency. Future research should extend the FRMLR model to dialectal variants, multimodal datasets, and other agglutinative and low-resource languages.