jo11223/tigrigna-stt

A fine-tuned Whisper model designed to accurately transcribe Tigrigna speech into text, optimized for the language’s unique sounds, vocabulary, and pronunciation patterns. # **Tigrigna Speech-to-Text Model** Unlock the power of **voice in Tigrigna**! This project fine-tunes **OpenAI Whisper** using **LoRA (Low-Rank Adaptation)** to transcribe **Tigrigna speech into text** with high accuracy. Perfect for transcription, accessibility tools, and voice-driven applications. --- ## **✨ Key Features** * 🚀 Fine-tuned **Whisper-small** for Tigrigna * 💡 Uses **LoRA** for efficient GPU memory usage * 🗂️ Handles custom Tigrigna datasets with TSV + audio format * 📊 Tracks **training loss** and **Word Error Rate (WER)** * 🏁 Exports a **final LoRA-adapted model** for instant inference * 🎯 Provides **sample transcription examples** for quick testing --- ## **🛠️ Setup & Dependencies** Install required packages: ```bash !pip install git+github.com !pip install torchaudio librosa datasets jiwer evaluate !apt-get install -y ffmpeg ``` Additional libraries: * **Transformers (Hugging Face)** – tokenizer, feature extractor, Whisper model * **PEFT (LoRA)** – efficient fine-tuning * **Torch + CUDA** – GPU training * **Google Drive (Colab)** – store datasets and checkpoints * **Matplotlib, tqdm, pandas, numpy** – data handling & visualization --- ## **📂 Dataset Format** Your dataset should contain: * **TSV file** (`data.tsv`) with `path` & `sentence` columns * **Audio folder** (`recordings/`) containing `.wav` or `.flac` files Example TSV: ```text path sentence audio1.wav ሰላም ከመይ ኣሎ? audio2.wav መልእክት ላክልኝ። ``` Update paths in the script: ```python dataset_dir = 'your-dataset-path' shared_drive_dir = 'your-shared-drive-path' ``` --- ## **🏋️ Training Process** 1. Mount **Google Drive** to store checkpoints & final model 2. Prepare dataset using Hugging Face `Dataset` API 3. Tokenize audio-text pairs and extract features 4. Apply **LoRA configuration** for memory efficiency 5. Train for **8 epochs** using AdamW optimizer 6. Auto-save checkpoints after each epoch 7. Evaluate **WER** on test set 8. Visualize …