Offline Uyghur ASR Demo on Android
This repository hosts two Android applications for offline Uyghur Automatic Speech Recognition (ASR), built upon sherpa‑onnx and the SenseVoice Int8 ONNX model. In addition, it provides the SenseVoice Int8 ONNX model itself along with its corresponding vocabulary, which contains approximately 16,000 Uyghur tokens.
The project demonstrates the deployment of Uyghur ASR models on mobile devices, including an ASR evaluation application and a pseudo-streaming recognition demo.
🧠 Model & Training Details
The underlying acoustic model is built upon SenseVoice — a powerful multi-language speech foundation model. However, the original SenseVoice release does not include Uyghur (a low-resource Turkic language).
To bridge this gap, we fine-tuned the SenseVoice model using 118 hours of Uyghur speech data from the CommonVoice corpus. The model was trained for 100 epochs to adapt the representations to Uyghur phonetics and linguistics.
Key adaptations include:
- Vocabulary reconstruction to support Uyghur-specific characters and pronunciation rules.
- Efficient fine-tuning strategies to preserve the original model's generalization while achieving high accuracy on Uyghur.
📱 Applications
1. Uyghur ASR Evaluation APK
An Android application for evaluating Uyghur ASR models on real mobile devices.
Supported metrics:
- Word Error Rate (WER)
- Character Error Rate (CER)
- Real-Time Factor (RTF)
Users can load Uyghur speech recordings and obtain offline recognition results without network connection.
2. Uyghur Pseudo-Streaming ASR Demo APK
A real-time Uyghur ASR demonstration application.
Features:
- Microphone-based recognition
- Chunk-based pseudo-streaming inference
- Offline speech processing
- Low-latency text generation
The demo provides a voice assistant-like experience for Uyghur speech recognition on mobile devices.
🙏 Acknowledgements
We would like to express our sincere gratitude to the SenseVoice team (GitHub) for releasing their powerful multi-language speech foundation model, which serves as the backbone of our fine-tuning work. Their open-source efforts have greatly accelerated research in low-resource speech recognition.
We also extend our heartfelt thanks to the sherpa-onnx developers (GitHub) for providing an efficient and lightweight ONNX inference framework that enables seamless on-device deployment with minimal latency. Their well-designed toolkit makes it possible to bring complex ASR models to mobile platforms effortlessly.
This project would not have been possible without the outstanding contributions of these open-source communities. We are grateful for their dedication and hope our work can in turn benefit the broader ecosystem.
📦 Deployment
Both applications are built with sherpa-onnx and support ONNX-based Uyghur ASR models.
The whole pipeline runs locally on Android devices without uploading audio data to external servers.
This project provides a practical solution for deploying and evaluating lightweight Uyghur ASR systems on edge devices.
🗺️ Open-Source Roadmap
We are committed to full transparency and reproducibility. Our open-source plan is phased as follows:
| Phase | Status | Deliverables |
|---|---|---|
| Phase 1 | ✅ Released | Android APKs (Evaluation & Pseudo-streaming Demo) |
| Phase 2 | ✅ Released | Exported ONNX weights (uyghur vocabsize=16000) |
| Phase 3 | 🔜 Coming Soon | PyTorch (PT) weights (original fine-tuned checkpoint) |
Stay tuned for updates! We aim to empower the low-resource NLP community by sharing both the models and the complete methodology.
📬 Citation & Contact
For any questions or collaboration, feel free to open an issue on this repository.