abrahamkifleT/amharic-tts-stt

# Amharic Text-to-Speech & Speech-to-Text (From Scratch) A research-grade, fully open-source pipeline for training and serving **Amharic (አማርኛ)** speech AI models — built from scratch using PyTorch, no external APIs. --- ## Features - 🧠 **ASR (Speech → Text):** Conformer + CTC decoder trained on Amharic audio - 🔊 **TTS (Text → Speech):** Tacotron2 acoustic model + HiFi-GAN vocoder - 📝 **Amharic G2P:** Custom Grapheme-to-Phoneme for all Ethiopic Unicode blocks - 🎙️ **Recording Studio:** Built-in web UI to record & contribute your own voice data - 📊 **Training Dashboard:** Live loss/WER visualizations in the browser - 🚀 **Colab Notebooks:** Ready-to-run GPU training notebooks (free GPU) --- ## Project Structure ``` ├── data/ # Datasets & processed data │ ├── raw/ # Downloaded raw speech datasets │ ├── processed/ # Preprocessed features (mel, tokens) │ └── recordings/ # Your own recorded clips ├── src/ │ ├── data/ # G2P, preprocessing, PyTorch datasets │ ├── models/ │ │ ├── asr/ # Conformer-CTC ASR model │ │ └── tts/ # Tacotron2 + HiFi-GAN TTS model │ ├── training/ # Training loops & metrics │ └── inference/ # Inference scripts ├── notebooks/ # Google Colab training notebooks ├── checkpoints/ # Saved model weights ├── app/ # Flask web inference app │ ├── server.py │ ├── templates/ │ └── static/ ├── requirements.txt └── README.md ``` --- ## Quick Start ### 1. Install Dependencies ```bash python -m venv venv venv\Scripts\activate # Windows pip install -r requirements.txt ``` ### 2. Download Datasets ```bash python data/download_datasets.py --datasets fleurs common_voice alffa ``` ### 3. Preprocess Data ```bash # For ASR python src/data/preprocess_asr.py --input data/raw --output data/processed/asr # For TTS python src/data/preprocess_tts.py --input data/raw --output d …