Speech to text app for Nigerian accents

# Nigerian Accented Speech-to-Text Model A fine-tuned version of OpenAI's Whisper model, trained to better understand Nigerian-accented English speech. ## What It Does Most speech-to-text (STT) systems struggle with Nigerian accents because they're mainly trained on American or British English. This project fine-tunes Whisper on Nigerian-accented voice data to improve transcription accuracy for Nigerian speakers. ## How It Was Built - **Base model:** `openai/whisper-small` - **Training data:** Nigerian-accent subset of the Mozilla Common Voice dataset (`benjaminogbonna/nigerian_common_voice_dataset`) - **Language:** Python - **Core libraries:** - `torch` (PyTorch) — model training, tensor operations, GPU handling - `transformers` (Hugging Face) — Whisper model, tokenizer, and feature extractor - `datasets` (Hugging Face) — loading and splitting the training data - `huggingface_hub` — authenticating and downloading models/datasets - **Evaluation:** - `evaluate` — scoring model outputs - `jiwer` — calculating Word Error Rate (WER) and Character Error Rate (CER) - `rapidfuzz` — string alignment used internally by `jiwer` ## Why It Matters Accurate Nigerian-accented STT can support: - Medical documentation (transcribing doctor-patient conversations) - Customer service and voice assistants - Call analysis for sales/support quality - Video captioning and searchable media - Voice dictation in apps and internal tools ## What's Next The model currently only covers Nigerian-accented **English**. Future work will expand it to **Nigerian Pidgin, Yoruba, and Igbo**, to make speech-to-text technology more inclusive of Nigeria's language diversity.