Wakeword Detection Android App

This Android application performs real-time wake word detection using ONNX models. The system integrates custom audio preprocessing, Mel-spectrogram generation, feature embedding, and wake word classification — all directly on-device. The backend is fully implemented in Java, optimized for Android runtime environments.

Features

• Real-Time Wake Word Detection: Detects custom wake words from live microphone input.
• ONNX Runtime Integration: Runs ONNX-based inference for Mel-spectrogram, embedding, and classification models.
• Custom Pipeline: Uses a modular Java pipeline: audio buffer → Mel-spectrogram → embedding → wake word prediction.
• YAMNet Classification: Optionally integrates TensorFlow Lite YAMNet to classify ambient audio.
• Thread-Safe Design: Handles audio recording, inference, and UI updates with optimized multithreaded architecture.
• Modular & Extensible: Easily replace models or adjust buffer/window configurations for different use cases.

Architecture Overview

AudioRecord (16 kHz PCM)
       ↓
1280-sample sliding buffer
       ↓
Mel-spectrogram (ONNX)
       ↓
Embedding model (ONNX)
       ↓
Wake word classifier (ONNX)
       ↓
Detection result (UI update via callback)

Prerequisites

• Android Studio (latest stable version)
• Android device running Android 6.0 (API 23) or higher
• Internet access for downloading dependencies (e.g., ONNX Runtime Mobile)
• Place the following models in your assets/ folder:
  • melspectrogram.onnx
  • embedding_model.onnx
  • your_custom_wakeword_model.onnx (e.g., chaamiiya.onnx)
  • yamnet.tflite and yamnet_labels.txt for sound classification

Usage

1. Clone this repository.
2. Open the project in Android Studio.
3. Place the required .onnx and .tflite models in the assets/ folder.
4. Build and run the app on a real device or emulator (microphone access required).
5. Speak the wake word and observe detection results in the UI.

Notes

• This project provides a fully on-device wake word detection pipeline using ONNX models — no server or internet required after setup.
• The wake word model and embedding extractor are fully customizable; you can retrain models using Python and deploy them here.