As part of the thesis, an Android application for wake word recognition is implemented. Recognition is performed using a locally stored TensorFlow Lite model on the device. The model is trained using MFCCs obtained from a custom set of audio recordings.
The application operates by initially capturing audio from the device's input, subsequently transforming it into features, and then conducting classification on the resulting matrix. This process enables us to achieve continuous word recognition. The processing in the application must be equivalent to the processing from the model training. The model achieves an accuracy of 88.73% on test data, while the application, based on user testing, is 82.23% accurate on real-world data.
|
