Short Utterance Based Speaker Identification System For Resource Constrained Devices
Sanghamitra V. Arora, Rekha Vig · 2018
Device manufacturers and developers are seeking viable options to enable users to interact with their wearable devices, which have a limited graphical user interface. One of the upcoming technologies in this sector is a voice user interface. But there are challenges to implementing voice-based identification methods on wearable devices due to the constrained resources. To overcome the disadvantages of short speech utterance, hybridized speech production features are proposed-Linear predictive coefficient (LPC) feature, Shifted delta cepstral (SDC) feature, Cochlea gram (CG) features, Pitch related features. A simplified version of KNN approach is used in order to classify the speakers. Overall performance of TIMIT database (which is text independent) is shown to give 97.3% recognition accuracy. The method can be useful for speaker authentication in rural areas since there would be no language barrier.