Master's Thesis presentation, Nils Hammaräng Grip and Elias Nilsson

Abstrakt:

The presentation will go through the creation of a keyword spotting (KWS) system which aims to find specific keywords within speech. The system is based on feature extraction (from sound waves captured by a microphone) and machine learning. Several different methods have been considered throughout the project. Unlike many previous studies, this work has focused on creating a system which can potentially be used in real-time, and more specifically in an automotive environment. Previous results, within KWS, and findings relevant for KWS within an automotive environment have thus been taken into consideration. Another important aspect, in this project, is the complexity of the model, which will appear as a time latency for the user. Therefore, both performance and complexity are taken into consideration when evaluating the models. The performance is examined using recorded speech and also in real-time, both in a noise-free environment and a noisy automotive environment. The results are discussed, and based on a performance-complexity trade-off analysis, a final choice of model is made.

Examiner: Serik Sagitov
Supervisor: Amir Eghbali (APTIV)