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.

Signal space diversity (SSD) introduced by Boutros and Viterbo tremendously improves the error performance over fading channels without any power or bandwidth sacrifice. Maximum benefit is obtained with full diversity (FD) multidimensional codebooks and the rather complex maximum likelihood (ML) detection. In the present work, we propose a generalized combinatorial representation and an associated low-complexity list-based detection algorithm that works for any non-full or full diversity multidimensional codebook. The algorithm includes three options of lists offering different complexity-performance trade-offs. One of the lists has random size and is guaranteed to contain the ML point, while the other two lists have fixed or limited size. Some analytical results are presented for the list sizes. The algorithm also includes a smart search to find the most probable codebook point in the list. Numerical results show that the proposed algorithm yields optimal or close-to-optimum performance with a noteworthy detection complexity reduction. Therefore, its adoption can be of practical interest, especially in applications where powerful error-correcting codes are not supported.

Previous transmission techniques for fading channels that exploit signal space diversity (SSD) are based on rotation or precoding of an original signal constellation. In this work, we follow a completely different approach to design full diversity multidimensional codebooks whose implementation has low complexity and requires little storage. The design consists of finding good permutations of the labels of points in a fixed grid. We consider both the uniform and the Gaussian-shaped grids. An ensemble average performance based on an enumerative analysis of the error events is presented, which brings some insight into the problem. Three low-complexity algorithms are then proposed to find good permutations sets. Both analytical and simulation results confirm the good performance of the proposed codebooks, which can be considered as good candidates for practical communications systems.

In this paper, the use of network coding is exploited to improve the cooperative diversity order of multi-source multi-hop networks with line topology. We elaborate on a previous system model by allowing user nodes to be arbitrarily located along the line network, to have the ability to perform network decoding, and to be able to listen to upstream nodes. The recently introduced generalized dynamic-network codes (GDNC) are shown to be the best performing scheme under several distinct circumstances. Computer simulation results support the superiority of the GDNC scheme, and suggest that the connection between network coding and classical channel coding, initiated in a previous work, can once again be proved useful by reducing the design of network codes for arbitrarily located nodes to the design of classical block codes with unequal error/erasure protection.