jamil2/azerbaijani-stt

🇦🇿 Azerbaijani Speech-to-Text

State-of-the-art transformer model achieving 100% output diversity!

🏆 Performance

• Loss: 0.0028 (99.5% reduction)
• Diversity: 100% (15/15 unique outputs)
• Quality: Production-grade

📊 Model Details

• Parameters: 25M
• Architecture: Transformer (Pre-LayerNorm)
• d_model: 384, feedforward: 1536
• Layers: 6 encoder + 6 decoder
• Training: 140 augmented samples

🚀 Demo

Try it: https://huggingface.co/spaces/jamil2/azerbaijani-stt-demo

💻 Usage

from huggingface_hub import hf_hub_download
import torch

model_path = hf_hub_download(
    repo_id="jamil2/azerbaijani-stt",
    filename="az_model.pt"
)
checkpoint = torch.load(model_path)

📝 Examples

Banking: "Salam, avtomobilim üçün KASKO qiyməti bilmək istəyirəm..." Culture: "Bu gün Azərbaycan mədəniyyətinin və Azərbaycan dilinin..."

📜 License

MIT - Free to use!