🖊️ IndicSwipe v2

A state-of-the-art gesture typing (Swipe-to-Type) app tailored for Indian languages. IndicSwipe combines a deep character-level Transformer gesture decoder with n-gram language models running on-device for fast, accurate, and fluid typing across 22+ Indian languages and romanized dialects (Hinglish, Tanglish, etc.).

---

📺 Overview & Demo

---

🌐 Supported Languages & Downloads

IndicSwipe supports 22 scheduled languages of India and their corresponding keyboard layouts, along with their romanized variants. You can download the pre-trained models and dataset packages directly from the table below:

Language	Script / ISO Code	Accent Color	Transliteration	Swipe Decoding
Hindi	__hi__ / हिन्दी		📥 Download	📥 Download
Bengali	__bn__ / বাংলা		📥 Download	📥 Download
Tamil	__ta__ / தமிழ்		📥 Download	📥 Download
Telugu	__te__ / తెలుగు		📥 Download	📥 Download
Marathi	__mr__ / मराठी		📥 Download	📥 Download
Kannada	__kn__ / ಕನ್ನಡ		📥 Download	📥 Download
Gujarati	__gu__ / ગુજરાતી		📥 Download	📥 Download
Punjabi	__pa__ / ਪੰਜਾਬੀ		📥 Download	📥 Download
Malayalam	__ml__ / മലയാളം		📥 Download	📥 Download
Odia	__or__ / ଓଡ଼ିଆ		📥 Download	📥 Download
Assamese	__as__ / অসমীয়া		📥 Download	📥 Download
Maithili	__mai__ / मैथिली		📥 Download	📥 Download
Sanskrit	__sa__ / संस्कृतम्		📥 Download	📥 Download
Urdu	__ur__ / اردو		📥 Download	📥 Download
Kashmiri	__ks__ / کٲشُر		📥 Download	📥 Download
Nepali	__ne__ / नेपाली		📥 Download	📥 Download
Sindhi (Arabic)	__sd__ / سنڌي		📥 Download	📥 Download
Sindhi (Devanagari)	__sdd__ / सिंधी		📥 Download	📥 Download
Konkani	__gom__ / कोंकणी		📥 Download	📥 Download
Manipuri	__mni__ / ꯃꯩꯇꯩꯂꯣꯟ		📥 Download	📥 Download
Bodo	__brx__ / बड़ो		📥 Download	📥 Download
Dogri	__doi__ / डोगरी		📥 Download	📥 Download
Santali	__sat__ / ᱥᱟᱱᱛᱟᱲᱤ		📥 Download	📥 Download

---

🛠️ Model Training & Pipeline

The pipeline is split into modular scripts located in the training/ directory.

Step 1: Generate Gesture Datasets

Generate synthetic swipe trajectory datasets based on keyboard geometry and target lexicons.

python training/01_generate_dataset.py --lang marathi --output training/data/marathi_train.jsonl

Step 2: Train the Transformer Model

Train the CharacterLevelSwipeModel with standard trajectory features (normalized coordinates, velocity, acceleration, nearest key embeddings) and character tokenizer.

python training/02_train_model.py \
    --lang marathi \
    --train_path training/data/marathi_train.jsonl \
    --val_path training/data/marathi_val.jsonl \
    --checkpoint_dir checkpoints/marathi/

Step 3: Export PyTorch to ONNX

Compile the PyTorch model encoder and decoder to standard ONNX with dynamic shapes.

python training/03_export_onnx.py \
    --checkpoint checkpoints/marathi/best_model.pt \
    --out_dir android_ready/marathi/

Step 4: Graph Optimization

Perform ONNX constant folding, dead-node elimination, and sequence fusing.

python training/04_optimize_onnx.py \
    --model_dir android_ready/marathi/

Step 5: Post-Training Quantization

Quantize the weights to 8-bit integers (INT8) to reduce model size by 4x with negligible accuracy drop.

python training/05_quantize_onnx.py \
    --model_dir android_ready/marathi/

---

📱 Android App Integration

The Android Keyboard App located in indic_swipe_android_studio_bundle_1 features:

• ONNX Runtime Mobile: Integrated local inference using ONNXRuntime to execute the gesture decoder.
• KenLM JNI: Custom C++ bindings compiled using NDK (kenlm-jni) to perform blazing-fast N-Gram language model language scoring on the decoder's beam outputs.
• Swipe Trail Renderer: Premium visual physics-based smooth finger trails (SwipeView.kt).

To run the Android App:

1. Open the indic_swipe_android_studio_bundle_1 directory in Android Studio.
2. Build the C++ JNI libraries using NDK CMake.
3. Place your optimized models (swipe_model_character_quant.onnx and language files) in the app/src/main/assets directory.
4. Build and run on an Android Device/Emulator!

---

👥 Contributors

• Srihari S
• Thanmay Jayakumar
• Raj Dabre

---

🤝 Acknowledgements

We would like to express our gratitude to the following individuals for their support (listed in alphabetical order):

• Deepon Halder
• Kaushal Bhogale
• Krishna Jeena
• Mohammed Safi Ur Rahman Khan
• Pranjal Chitale
• Sidharth Pulipaka
• Tahir Javed
• Vignesh Selvaraj

---

📚 References

1. IndicSwipe:
   • Paper: "Joint Transformer/RNN Architecture for Gesture Typing in Indic Languages" (COLING 2020)
   • Repository: AI4Bharat/Indic-Swipe
2. Aksharantar:
   • Paper: "Aksharantar: Towards Building Open Transliteration Tools for the Next Billion Users" (EMNLP 2023)
   • Repository: AI4Bharat/IndicXlit
   • Dataset: ai4bharat/Aksharantar on Hugging Face
3. CleverKeys:
   • Repository: tribixbite/CleverKeys
4. Google Keyboard (Gboard) Gesture Typing:
   • Paper: "Long Short Term Memory Neural Network for Keyboard Gesture Decoding" (Google Research)
   • Link: Google Research Publication PDF
5. Google Keyboard Blog:
   • Blog Post: "The Machine Intelligence Behind Gboard"
   • Link: Google Research Blog

---

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.