Abstract
Our work proposes a hardware architecture for a Long Short-Term Memory (LSTM) Neural Network, aiming to outperform software implementations, by exploiting its inherent parallelism. The main design decisions are presented, along with the proposed network architecture. A description of the main building blocks of the network is also presented. The network is synthesized for various sizes and platforms, and the performance results are presented and analyzed. Our synthesized network achieves a 251 times speed-up over a custom-built software network, running on an i7-3770k Desktop computer, proving the benefits of parallel computation for this kind of network.

Abstract
In this paper we address the transformation of whispered speech into natural voiced speech. Representative state-of-the-art solutions are first reviewed as well as a baseline algorithm. For the most part, these solutions fall in the realm of voice conversion strategies since the output signal is obtained as a projection of an input signal. In this paper, we propose a different approach that addresses flexible parametric synthesis of the voiced signal component, as well as its implantation on the whispered signal, in a linguistically consistent way and while trying to convey idiosyncratic information. The most critical functions of phonetic segmentation, spectral envelope estimation, arbitrary periodic wave shape synthesis, and F0 modulation, are described and their operation illustrated with examples.

Abstract
In recent years, tools in perceptual coding of high-quality audio have been tailored to capture highly detailed information regarding signal components so that they gained an intrinsic ability to represent audio parametrically. In a recent paper, we described a first validation model to such an approach applied to parametric coding of wideband speech. In this paper we describe specific advances to such an approach that improve coding efficiency and signal quality. A special focus is devoted to the fact that persistent transmission to the decoder of phase information is avoided, to the synthesis of both impulse-like and noise-based plosives using short-term windows, to improved ways of spectral envelope modelling, and to the fact that direct synthesis in the time-domain of the periodic content of speech is allowed in order to cope with fast F0 changes. A few examples of signal coding and transformation illustrate the impact of those improvements.

Abstract
Interference is a fundamental issue in wireless mesh networks (WMNs) and it seriously affects the network performance. In this paper we characterize the power interference in IEEE 802.11 CSMA/CA based wireless mesh networks using directional antennas. A model based centralized call admission control (CAC) scheme is proposed which uses physical collision constraints, and transmitter-side, receiver-side and when-idle protocol collision prevention constraints. The CAC assists to manage requests from users depending on the available bandwidth in the network: when a new virtual link establishment request from a user is accepted into the network, resources such as interface, bandwidth, transmission power and channel are allocated in the participating nodes and released once the session is completed. The proposed CAC is also able to contain the interference in the WMN by managing the transmission power of nodes.

Abstract
This paper describes the evaluation of a multi-hop wireless networking solution for Smart Grid metering in an industrial environment. The solution relies on RPL, 6LoWPAN, and IEEE 802.15.4g protocols, and has been implemented using low-power and low-capacity devices. Also, it supports both TCP and UDP protocols to transport traffic from DLMS/COSEM Smart Grid metering applications. The experimental tests took place in an industrial environment during 20 days. The obtained results allowed the characterization and evaluation of the developed solution and can be used as a basis to evaluate other 6LoWPAN/IEEE 802.15.4g networking solutions.

Abstract
Wireless Underground Networks (WUN) have many applications, such as border surveillance, agriculture monitoring, and infrastructure monitoring. Recent studies have shown that they are feasible and have deployment advantages over wired networks, but only a few WUN evaluations in multiple access scenarios have been done. This paper presents a simulation study on medium access for a WUN with 4 nodes buried, and one node aboveground. The simulations were carried out using the ns-3 simulator and they evaluate both Wi-Fi, and Lr-Wpan networks for dry and wet soils. We verified that for the same number of concurrent nodes, the use of the RTS/CTS mechanism has a much higher influence than the soil water content. Furthermore, a study about the feasibility of using Wi-Fi fingerprinting for positioning above the ground based on the buried infrastructure revealed promising results.