No abstract

Keyword spotting (KWS) is one of the speech recognition tasks most sensitive to the quality of the feature representation. However, the research on KWS has traditionally focused on new model topologies, putting little emphasis on other aspects like feature extraction. This paper investigates the use of the multitaper technique to create improved features for KWS. The experimental study is carried out for different test scenarios, windows and parameters, datasets, and neural networks commonly used in embedded KWS applications. Experiment results confirm the advantages of using the proposed improved features.

International audienceRecently, a time-frequency approach for testing stationarity was proposed. However, this method inefficiently detects nonstationarities of the first-order. Here, we present two contributions that improve the test performance and allow the detection of first-order evolutions. The first one is to use an adequate distance measure. The second is a modification of the method in order to consider the spectral content from the signal itself when computing the distances

No abstract

Acoustic Scene Classification (ASC) systems have great potential to transform existing embedded technologies. However, research on ASC has put little emphasis on solving the existing challenges in embedding ASC systems. In this paper, we focus on one of the problems associated with smaller ASC models: the generation of smaller yet highly informative training datasets. To achieve this goal, we propose to employ the so-called multitaper-reassignment technique to generate high-resolution spectrograms from audio signals. These sharp time-frequency (TF) representations are used as inputs to a splitting method based on TF-related entropy metrics. We show via simulations that the datasets created through the proposed segmentation can successfully be used to train small convolutional neural networks (CNNs), which could be employed in embedded ASC applications.