README.md

ASR Demo App

This app demonstrates running speech recognition models on Android using ExecuTorch. It supports both Whisper and Parakeet model families.

Supported Models

Model	Type	Details
Whisper Tiny/Small/Medium (INT8/INT4)	Streaming	Requires model, tokenizer, and preprocessor
Whisper Tiny/Small/Medium (FP32)	Streaming	Requires model, tokenizer, and preprocessor
Parakeet TDT 0.6B (INT4)	Synchronous	Requires model and tokenizer

Export Model Files

• Whisper: Follow the instructions at https://github.com/pytorch/executorch/tree/main/examples/models/whisper
• Parakeet: Follow the instructions at https://github.com/pytorch/executorch/tree/main/examples/models/parakeet

Run the App

1. Open AsrApp in Android Studio
2. Copy the executorch.aar library (with ASR and Parakeet JNI bindings) into app/libs/
3. Build and run on device

Download Models

The app includes a built-in download screen to fetch models from HuggingFace. Alternatively, push files manually:

adb push model.pte /data/local/tmp/asr/
adb push tokenizer.json /data/local/tmp/asr/
adb push whisper_preprocessor.pte /data/local/tmp/asr/  # Whisper only

Recording Behavior

• Whisper: Click to start recording; automatically stops after 30 seconds
• Parakeet: Click to start recording; click again to stop (no time limit)