Master Thesis Project @TUGraz, 01 / 2022 Author: Christian Walter

The aim of this project was to implement a Key-Word-Spotting (KWS) system with neural networks trained on the speech command dataset [1] and applied in a simple video game. Several low computational Convolutional Neural Networks (CNN) in the fashion of [2] were evaluated. A pre-training method using Generative Adversarial Networks (GAN) [3] was used to increase the accuracies. Wavenets [4] were extended with a classification of speech commands.

The gameplay video can be found in the folder "docs/gameplay.mp4"

Operating System: Linux Programming Language: Python >=3.8.x Neural Network Framework: Pytorch >=1.7 Several packages must be installed in order to run the game. Try to run the program and determine the missing dependencies. Further, a microphone is necessary to play the game.

To play the game you have to open a terminal in the project path and type: ´python kws_main.py´

[1] Warden, P., “Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition”, 2018. [2] Sainath, T., "Convolutional neural networks for small-footprint keyword spotting", 2015. [3] Goodfellow, I. J., “Generative Adversarial Networks”, 2014. [4] Oord, A., "WaveNet: {A} Generative Model for Raw Audio", 2016.