Estratégias de Combinação de Espectrogramas de Magnitude e de Fase Aplicadas em Sistemas Robustos de Detecção de Palavras-Chave
Ênio Silva, Rui Seara

DOI: 10.14209/sbrt.2021.1570726876
Evento: XXXIX Simpósio Brasileiro de Telecomunicações e Processamento de Sinais (SBrT2021)
Keywords: Stacked generalization Keyword spotting Phase spectrograms Feature extraction
Abstract
The demand for keyword spotting (KWS) systems has been increasing considerably for the most diverse applications in the real world. However, the performance of these systems is severely degraded under operating conditions with a low signal-to-noise ratio (SNR). Aiming to obtain KWS systems robust to noise, this research work presents an investigation about the feature extraction process performed in these systems. Particularly, the present work proposes the use of strategies for combining features taking into account the magnitude and phase spectrograms of speech signals. In this way, KWS systems using feature extraction considering combinations of the magnitude and phase are contrasted with those using only magnitude spectrograms. Numerical simulation results are shown and assessed with respect to the accuracy of keyword recognition, confirming the effectiveness of the strategies used in this work.

Download