Tor Andre Myrvoll

Optimal clustering and non-uniform allocation of Gaussian kernels in scalar dimension for HMM compression [speech recognition applications]

We propose an algorithm for optimal clustering and nonuniform allocation of Gaussian kernels in scalar (feature) dimension to compress complex, Gaussian mixture-based, continuous density HMMs into computationally efficient, small footprint models. The symmetric Kullback-Leibler divergence (KLD) is used as the universal distortion measure and it is minimized in both kernel clustering and allocation procedures. The algorithm was tested on the resource management (RM) database. The original context-dependent HMMs can be compressed to any resolution, measured by the total number of clustered scalar kernel components. Good trade-offs between the recognition performance and model complexities have been obtained; the HMM can be compressed to 15-20% of the original model size, which needs 1-5% of multiplication/division operations, and results in almost negligible recognition performance degradation.

Penalized Logistic Regression With HMM Log-Likelihood Regressors for Speech Recognition

Hidden Markov models (HMMs) are powerful generative models for sequential data that have been used in automatic speech recognition for more than two decades. Despite their popularity, HMMs make inaccurate assumptions about speech signals, thereby limiting the achievable performance of the conventional speech recognizer. Penalized logistic regression (PLR) is a well-founded discriminative classifier with long roots in the history of statistics. Its classification performance is often compared with that of the popular support vector machine (SVM). However, for speech classification, only limited success with PLR has been reported, partially due to the difficulty with sequential data. In this paper, we present an elegant way of incorporating HMMs in the PLR framework. This leads to a powerful discriminative classifier that naturally handles sequential data. In this approach, speech classification is done using affine combinations of HMM log-likelihoods. We believe that such combinations of HMMs lead to a more accurate classifier than the conventional HMM-based classifier. Unlike similar approaches, we jointly estimate the HMM parameters and the PLR parameters using a single training criterion. The extension to continuous speech recognition is done via rescoring of N-best lists or lattices.

On multiuser MIMO capacity benefits in air-to-ground communication for air traffic management

In this paper, we focus on communication from airplanes to Air traffic Control (ATC) towers at airports, employing multiple antennas at the ATC. The whole air-to-ground communication system can then be modeled as a MIMO MAC channel with mobile users. The main aim is to demonstrate that multiuser MIMO systems can be useful for achieving spectrally efficient aeronautical communications, to quantify the achievable capacity gains when using such systems, and to show that they can be implemented by deploying multiple antennas at reasonable inter-antenna distances at the ATC. In particular, we analyze the issue of antenna separation of a uniform linear array (ULA) at the ATC, for the case of line of sight (LoS) air-to-ground channels. It turns out that at high frequency (1 GHz), the use of large ULAs is indeed possible since the maximum (ergodic) capacity gain is achieved already for antenna separation on the order of a few centimeters. We also observe the system performance in Ricean fading channels. The performance is evaluated with respect to ergodic multiuser capacity throughout.

Measurement and modeling of the 5 GHz airport surface channel at Barajas Airport

AeroMACS is a system currently under development to be used for airport surface communications. It is based on the IEEE802.16–2009 standard, and is developed in cooperation with the WiMAX forum. The frequency band allocated to AeroMACS is 5091–5150 MHz. When specifying the AeroMACS system, knowledge of typical propagation conditions at airports is of importance. Typical path loss models are necessary to estimate the range of a transmitter within different zones of the airport, and typical fading characteristics are used to estimate the performance of e.g. the selected coding and modulation schemes.

User requirements for HEO SATCOM for ATM in high latitudes

It is currently a significant ongoing effort worldwide to develop the future Air Traffic Management (ATM) system. As part of this work, a satellite communication system may ease the congestion problem for ATM services in high density airspace, and in addition provide coverage in oceanic, remote and polar (ORP) areas. For coverage over polar areas, satellites in highly elliptical orbits (HEO) are particularly suitable. In this paper, an overview of user categories is given and the channel characteristics of an aeronautical satellite channel are considered. Both Molniya and Tundra orbits are included. Curves show how parameters like elevation angle, free space path loss and Doppler shift vary as function of satellite movements. In addition, atmospheric effects due to signal propagation through the ionosphere and the troposphere is considered, and finally the effect of multipath propagation due to signal reflections by the aircraft surface and ground.

When SDR meets a 5G candidate waveform : Agile use of fragmented spectrum and interference protection in PMR networks

Filter bank multi-carrier (FBMC) is a candidate modulation scheme for 5G cellular mobile broadband networks. A specific domain where FBMC offers a clear advantage over other multicarrier solutions is the efficient occupancy of underutilized and fragmented spectrum. This is due to the rich spectral containment of the FBMC technology, which guarantees superior interference protection to the primary coexisting transmissions. These features of FBMC could play an important role in the economic delivery of 5G services in licensed and unlicensed bands. This article focuses on the public safety domain where existing professional mobile radio (PMR) users plan to add broadband services at the 400 MHz band, aiming to occupy the spectral holes left by current narrowband transmissions. In this respect an agile SDR broadband downlink FBMC system, aimed at exploiting unused licensed PMR spectrum, was developed and experimentally validated. The FBMC frame structure shared key similarities with the LTE specification. Two different SDR design methodologies were used to build the real-time baseband prototype. The level of interference protection offered by a broadband FBMC system to a coexisting primary narrowband PMR transmission was practically demonstrated and compared to that of an equivalent LTE system. This was made feasible by evaluating the performance of a PMR terminal under different mobile channels and FBMC waveform configurations. The applicability of this work to other 5G spectrum cohabitation scenarios is discussed. Finally, the article highlights the need to extend the SDR design paradigm in order to tackle the challenges of real-time baseband processing in 5G broadband cellular systems.

Empirical Path Loss Models for C-Band Airport Surface Communications

AeroMACS is a system currently under development for airport surface communications. It is based on the IEEE802.16-2009 standard, and is developed in cooperation with the WiMAX forum. The frequency band allocated to AeroMACS is 5091-5150 MHz. When specifying the AeroMACS system, knowledge of typical propagation conditions at airports is of importance. Typically, path loss models are used to estimate the range of a transmitter within different zones of the airport, and typically fading characteristics are used to estimate the performance of, e.g., the selected coding and modulation schemes. Propagation measurements have been performed at the Barajas airport in Madrid, Spain. The measurements includes Non-Line-Of-Sight (NLOS) as well as Line-Of-Sight (LOS) conditions. Based on these measurements, path loss parameters and path loss models are developed for different zones of large airports.

N-Best Rescoring for Speech Recognition using Penalized Logistic Regression Machines with Garbage Class

State-of-the-art pattern recognition approaches like neural networks or kernel methods have only had limited success in speech recognition. The difficulties often encountered include the varying lengths of speech signals as well as how to deal with sequences of labels (e.g., digit strings) and unknown segmentation. In this paper we present a combined hidden Markov model (HMM) and penalized logistic regression machine (PLRM) approach to continuous speech recognition that can cope with both of these difficulties. The key ingredients of our approach are N-best rescoring and PLRM with garbage class. Experiments on the Aurora2 connected digits database show significant increase in recognition accuracy relative to a purely HMM-based system.

Minimum mean square error filtering of noisy cepstral coefficients with applications to ASR

In our previous work (2003), we investigated a new approach to robust speech recognition. An exact procedure was developed to filter noisy cepstral coefficients in the mean-square-error sense, and it was shown that this method outperformed the well known vector Taylor series (VTS) approach, which in turn is based on linear approximations to the non-linear filtering problem. Unfortunately. the procedure presented involved several integral equations with no known closed form solution. Numerical integration techniques were needed, which in turn led to slow performance, and in some cases, numerical problems. In this work we address this problem by using piecewise approximations to the integrands, which in turn yield closed form solutions. The revised procedure is tested on a subset of the Aurora 2 database, and the results are compared with the original numerical integration based approach, as well as VTS.

Dynamic Spectrum Access in realistic environments using reinforcement learning

We study the use of reinforcement learning to model Dynamic Spectrum Access in a realistic multi-channel environment. Three different approaches from the literature on the multi-armed bandit problem are compared on a set of realistic channel access models - two are based on stochastic models of the channel occupancy, while a third assumes an adversarial model. The algorithms are experimentally tested on channels occupied by primary users that behave according to a simple fair scheduler and a semi-Markov model based on WLAN traffic measurements; models that generate more realistic channel occupancy patterns than allowed by fixed i.i.d. probability models. The experiments show that the UCB1 algorithm of Auer et. al. [1] outperforms the other algorithms, and we support these findings using some simple theoretical results.