EVALUASI MODEL WAVENET UNTUK KLASIFIKASI PERINTAH SUARA REAL-TIME PADA KONTROL KARAKTER GAME

Abstract

Traditional game control systems such as keyboards and mice have limitations in providing an immersive and accessible experience. This research aims to develop a real-time voice control system for game character movement using the WaveNet deep learning architecture. The method used was training a WaveNetClassifier model on a voice command dataset consisting of four classes: "down", "up", "right", and "left". To improve resilience, the training data was augmented with background noise. The model was evaluated in three experimental runs (using seeds 42, 123, and 456) based on accuracy, precision, recall, F1-score, and latency metrics in two scenarios: clean test data (Hening) and noisy test data (Noise). The results show the WaveNetClassifier model is highly effective. The best performance (Seed 456) achieved 83.89% accuracy on clean data and 81.67% on noisy data. On average, the model's accuracy was very robust, with a performance drop of less than 0.4% when tested with noise. Furthermore, the model proved to be extremely fast with an average latency of only 5.20 ms, proving its feasibility for real-time gaming applications. The main weakness identified was the model's tendency to misclassify the "Left" command.