NKEthio/Amharic-Transcript-Generator

# Amharic Transcript Generator End-to-end Amharic speech-to-text pipeline for fine-tuning foundation models (FMs) on Amharic voice data and generating transcripts. ## What this project includes - Data loading from CSV manifests (`audio_path`, `transcript`) - Fine-tuning workflow for `openai/whisper-small` - Word error rate (WER) evaluation during training - Inference script for generating Amharic transcripts from audio - Config-driven training using YAML ## Project structure ```text configs/ amharic_whisper_ft.yaml scripts/ train_amharic_asr.py transcribe_audio.py src/ amharic_asr/ config.py data.py train.py transcribe.py tests/ test_config.py ``` ## Installation 1. Create and activate a Python 3.10+ virtual environment. 2. Install dependencies: ```bash pip install -r requirements.txt ``` ## Dataset format Prepare two CSV files: - `train.csv` - `validation.csv` Required columns: - `audio_path`: absolute or relative path to `.wav/.mp3/...` audio file - `transcript`: Amharic reference text Example: ```csv audio_path,transcript data/audio/sample1.wav,ሰላም እንዴት ነህ data/audio/sample2.wav,ይህ የአማርኛ ድምጽ ነው ``` ## Fine-tuning an FM for Amharic ASR 1. Edit `configs/amharic_whisper_ft.yaml` with your dataset paths and training settings. 2. Run: ```bash python scripts/train_amharic_asr.py --config configs/amharic_whisper_ft.yaml ``` Model checkpoints and final artifacts are saved to `outputs/amharic-whisper-small-ft` by default. Set `preprocessing_num_proc` in config to use more CPU cores during feature preparation. ## Generate transcript from audio ```bash python scripts/transcribe_audio.py \ --model-dir outputs/amharic-whisper-small-ft \ --audio-path data/audio/sample1.wav \ --chunk-length-s 30 ``` ## Notes - This repository provides the full mechanism (data -> fine-tune -> evaluate -> inference) for Amharic transcript generation. - You can switch to another FM by changing `base_model` in the config. ## Run on Google Colab You can run this project ea …