JONAHKYAGABA/-Kinyarwanda-ASR-Qualitative-Evaluation

Kinyarwanda ASR Qualitative Evaluation # 🎧 Kinyarwanda ASR Qualitative Evaluation This project performs **qualitative analysis** of an **Automatic Speech Recognition (ASR)** model on the **Kinyarwanda** language using Hugging Face pipelines. It compares transcriptions from the model against ground truth, highlights word-level differences, and displays them in color-coded format for easy human inspection. ## ✅ Features - Loads audio samples and reference transcripts from a dataset on Hugging Face - Performs transcription using a pre-trained ASR model (e.g. Whisper, MMS) - Computes WER and CER for each sample - Prints color-coded visual diffs between predicted and reference transcriptions: - ✅ Matches - ❌ Substitutions / Errors - Saves detailed CSV results per audio sample ## 🔤 Language Focus - **Language:** Kinyarwanda (`kin`) - **Dataset:** `KYAGABA/kinyarwanda_cleaned_testset` ## 🧪 Model Supports any Hugging Face ASR model, example: ```python model_name = "asr-africa/whisper-small-kinyarwanda_300hrs"