thetruesammyjay/speech-to-text-nigerian

A hybrid automatic speech recognition system leveraging Support Vector Machines and neural architectures for low-resource Nigerian languages including Igbo, Yoruba, Hausa, and Nigerian Pidgin. # Speech-to-Text System for Nigerian Languages A hybrid automatic speech recognition system leveraging Support Vector Machines and neural architectures for low-resource Nigerian languages including Igbo, Yoruba, Hausa, and Nigerian Pidgin. ## Project Overview This research project addresses the critical gap in speech recognition technology for Nigerian languages, where conventional deep learning approaches face significant limitations due to insufficient training data. By employing a hybrid architecture that combines the efficiency of Support Vector Machines with selective neural network components, this system delivers practical speech recognition capabilities optimized for resource-constrained environments. ### Research Context While modern ASR systems predominantly utilize Transformer-based architectures and end-to-end deep learning models trained on massive datasets, these approaches remain inaccessible for most African languages. The NaijaVoices dataset, comprising 1,867 hours of speech data across three major Nigerian languages, represents a substantial advancement but remains modest compared to the tens of thousands of hours available for high-resource languages like English or Mandarin. This project explores whether classical machine learning approaches, specifically Support Vector Machines, can deliver viable performance in low-resource scenarios while offering distinct advantages in computational efficiency, interpretability, and deployment feasibility on edge devices. ## Technical Architecture ### Core Components **Feature Extraction Pipeline** - Mel-Frequency Cepstral Coefficients (MFCC) extraction with 13-coefficient vectors - Pitch and fundamental frequency analysis for tonal language processing - Delta and delta-delta features for capturing temporal dynamics - Voice Activity Detection (VAD) for preprocessing optimization **Classification Framework** - Multi-class SVM with Radial Basis Function (RBF) kernels - One-vs-Rest strategy for phoneme …