njerikahoro/Whisper-Kikuyu

# Fine-Tuning Whisper for Kikuyu language Fine-tuning OpenAI's Whisper (small) for automatic speech recognition (ASR) in **Kikuyu (Gĩkũyũ)**, a low-resource Bantu language spoken in Kenya. This repository contains the training notebook, inference examples, and a Gradio demo interface used to fine-tune and evaluate the model. The resulting model is published at NjeriKahoro/Kikuyu-Whisper. ## Overview - **Base model:** `openai/whisper-small` - **Task:** Automatic Speech Recognition (transcribe) - **Language:** Kikuyu (`ki`) - **Training dataset:** `Anv-ke/kikuyu` (streamed, `Audio(decode=False)`) - **Framework:** 🤗 Transformers `Seq2SeqTrainer` - **Environment:** Google Colab (GPU runtime), Google Drive for checkpoint storage ## Repository Contents | File | Description | |------|--------------| | `Whisper.ipynb` | End-to-end notebook: dependency install, dataset loading & preprocessing, model fine-tuning, inference, and Gradio demo | ## Setup ```bash pip install "datasets **Note:** Inference currently uses `language="swahili"` as a proxy language tag, since Whisper does not have a native Kikuyu language token. This is a known workaround — see Limitations below. ## Interactive Demo The notebook includes a Gradio interface for live transcription and WER scoring against a reference transcript: ## Limitations - Whisper has no native Kikuyu language token, so `swahili` is used as the closest available proxy during decoding — this may introduce systematic biases in transcription. - The dataset is streamed rather than fully downloaded, so some corrupted or malformed audio samples are silently skipped during preprocessing. ## Related - Model card: NjeriKahoro/Kikuyu-Whisper - Demo Space: Omnilingual-Translation-Speech-RecognitionStudio - Dataset: Anv-ke/kikuyu