Publications

Gaussian mixture models (GMMs) have proven extremely successful for text-independent speaker recognition. The standard training method for GMMmodels is to use MAP adaptation of the means of the mixture components based on speech from a target speaker. Recent methods in compensation for speaker and channel variability have proposed the idea of stacking the means of the GMM model to form a GMM mean supervector. We examine the idea of using the GMM supervector in a support vector machine (SVM) classifier. We propose two new SVM kernels based on distance metrics between GMM models. We show that these SVM kernels produce excellent classification accuracy in a NIST speaker recognition evaluation task.

An active phased array radar with distributed low-peak-power transmit modules requires pulse compression to provide high sensitivity and fine range resolution. A long transmitted pulse, however, has accompanying problems such as a near-range blind zone for the transmitting channel and a loss of other gate data (dead gates) in other channels for a multichannel system. In this report the trade-off between the benefits and costs of pulse compression (lengthening) for multifunction phased array radars (MPARs) are analyzed. Specific results are presented for a three-channel MPAR and a two-channel terminal-area MPAR (TMPAR) that have been proposed as replacement systems for current U.S. civil-sector aircraft anad weather surveillance radar systems. The recommended maximum compression ratio is 130 ofr the MPAR and 80 for the TMPAR. The results are independent of radar peak power and antenna gain, and represent upper bounds. Acutal pulse compression ratios that would be employed are likely to be somewhat less tha these values, based on fulfilling specific sensitivity and scan-time requirements with specific radar physical parameters.

Support vector machines (SVMs) have proven to be a powerful technique for pattern classification. SVMs map inputs into a high-dimensional space and then separate classes with a hyperplane. A critical aspect of using SVMs successfully is the design of the inner product, the kernel, induced by the high dimensional mapping. We consider the application of SVMs to speaker and language recognition. A key part of our approach is the use of a kernel that compares sequences of feature vectors and produces a measure of similarity. Our sequence kernel is based upon generalized linear discriminants. We show that this strategy has several important properties. First, the kernel uses an explicit expansion into SVM feature space - this property makes it possible to collapse all support vectors into a single model vector and have low computational complexity. Second, the SVM builds upon a simpler mean-squared error classifier to produce a more accurate system. Finally, the system is competitive and complimentary to other approaches, such as Gaussian mixture models (GMMs). We give results for the 2003 NIST speaker and language evaluations of the system and also show fusion with the traditional GMM approach.

We consider the phenomenon of afterpulsing in avalanche photodiodes (APDs) operating in gated and free-running Geiger mode. An operational model of afterpulsing and other noise characteristics of APDs predicts the noise behavior observed in the free-running mode. We also use gated-mode data to investigate possible sources of afterpulsing in these devices. For 30-um-diam, 1.06-um-wavelength InGaAsP/InP APDs operated at 290 K and 4 V overbias, we obtained a dominant trap lifetime of td=0.32 us, a trap energy of 0.11 eV, and a baseline dark count rate 245 kHz.

The intelligibility of speech transmitted through low-rate coders is severely degraded when high levels of acoustic noise are present in the acoustic environment. Recent advances in nonacoustic sensors, including microwave radar, skin vibration, and bone conduction sensors, provide the exciting possibility of both glottal excitation and, more generally, vocal tract measurements that are relatively immune to acoustic disturbances and can supplement the acoustic speech waveform. We are currently investigating methods of combining the output of these sensors for use in low-rate encoding according to their capability in representing specific speech characteristics in different frequency bands. Nonacoustic sensors have the ability to reveal certain speech attributes lost in the noisy acoustic signal; for example, low-energy consonant voice bars, nasality, and glottalized excitation. By fusing nonacoustic low-frequency and pitch content with acoustic-microphone content, we have achieved significant intelligibility performance gains using the DRT across a variety of environments over the government standard 2400-bps MELPe coder. By fusing quantized high-band 4-to-8-kHz speech, requiring only an additional 116 bps, we obtain further DRT performance gains by exploiting the ear's insensitivity to fine spectral detail in this frequency region.

We have developed focal-plane arrays and laser-radar (ladar) imaging systems based on Geiger-mode avalanche photodiodes (APDs) integrated with high-speed all-digital CMOS timing circuits. A Geiger-mode APD produces a digital pulse upon detection of a single photon. This pulse is used to stop a fast digital counter in the pixel circuit, thereby measuring photon arrival time. This "photon-to-digital conversion" yields quantum-limited sensitivity and noiseless readout, enabling high-performance ladar systems. Previously reported focal planes, based on bump bonding or epoxy bonding the APDs to foundry chips, had coarse (100um) pixel spacing and 0.5ns timing quantization.

The optimization of traffic flows in highly congested airspace with rapidly varying convective weather is an extremely complex problem. Aviation weather systems such as the Corridor Integrated Weather System (CIWS) provide weather products and forecasts that aid en route traffic managers in making tactical routing decisions in convective weather, but traffic managers need automated decision support systems that integrate flight information, trajectory models and convective weather products to assist in developing and executing convective weather mitigation plans. A key element of an integrated ATM/wx decision support system is the ability to predict automatically when pilots in en route airspace will choose to deviate around convective weather and how far they will deviate from their planned path. The FAA Aeronautical Information Manual suggests that pilots avoid thunderstorms characterized by intense radar echo in en route airspace by at least 20 nautical miles (40 km). However, a recent study (Rhoda, et. al., 2002) of pilot behavior in both terminal and en route airspace near Memphis, TN suggested that pilots fly over high reflectivity cells in en route airspace and penetrate lower cells whose reflectivity is less than VIP level 3. Recent operational experience with CIWS supports the Rhoda findings (Robinson, et. al., 2004). This study presents initial results of research to develop a quantitative model that would predict when a pilot will deviate around convective weather in en route airspace. It also presents statistics that characterize hazard avoidance distances and weather penetrations. The results are based on the analysis of more than 800 flight trajectories through two Air Traffic Control (ATC) en route super-sectors (geographical regions that include several adjacent ATC en route sectors) on five days in the summer of 2003. One supersector from the Indianapolis Air Route Traffic Control Center (ZID ARTCC) encompassed southern Indiana, southwestern Ohio and northern Kentucky (ZID); the other, located in the Cleveland ARTCC (ZOB), included northern Ohio, along the southern shore of Lake Erie (ZOB). The weather encountered along the flight trajectories was characterized by the CIWS high-resolution precipitation (VIL) and radar echo tops mosaic (Klingle-Wilson and Evans, 2005) and NLDN lightning products. Flight trajectories were taken from the Enhanced Traffic Management System (ETMS).

We discuss two specific short term aviation weather forecasts - convection and ceiling - to illustrate the issues that arise in thinking about the overall decision support system, key users, and training needed to generate benefits. We also consider reducing weather-related fatal accidents. Second, what is the preexisting "baseline" of aviation forecasts/decision processes that already exists to address the user needs? In most cases, there are already various weather information sources that can be viewed as providing a short term forecast (e.g., a Center Weather Service Unit (CWSU) meteorologist, persistence, or animation loops of the past weather). How well do we understand how the "baseline" forecast and the associated user decision support system operate? How will the new forecast and its decision support compare? What are the training implications if the new forecast is rather different than the "baseline"? Third, how will we measure the change in system performance? For example, if the new forecast claims to help reduce delays and/or accidents, how will one address differences in the weather between the "before" and "after" time periods? How will one determine whether the new forecast is in fact the key factor, if there was a change? The paper concludes with some suggestions for development and testing of new aviation forecasts to improve safety and reduce delays.

The local airspace surrounding the San Francisco International Airport (SFO) is prone to regular occurrences of low ceiling conditions from May through October due to the intrusion of marine stratus along the Pacific coast. The low cloud conditions prohibit dual parallel landings of aircraft to the airport's closely spaced parallel runways, thus effectively reducing the arrival capacity by a factor of two. The behavior of marine stratus evolves on a daily cycle, filling the San Francisco Bay region overnight, and dissipating during the morning. Often the low ceiling conditions persist throughout the morning hours and interfere with the high rate of air traffic scheduled into SFO from mid-morning to early afternoon. The result is a substantial number of delayed flights into the airport and a negative impact on the National Air Space (NAS). Air traffic managers face a continual challenge of anticipating available operating capacity so that the demand of incoming planes can be metered to match the availability of arrival slots.

Air traffic congestion in the United States (US) National Airspace System (NAS) has increased significantly in the past ten years. This congestion has resulted in a rise of air traffic delays, which can cause massive monetary and human costs. When convective weather impacts jet routes and airport terminals, particularly within the most congested airspace sectors, it causes a reduction in traffic capacity that can lead to significant delays. In an effort to increase airspace capacity and reduce air traffic delays, the Massachusetts Institute of Technology Lincoln Laboratory (MIT LL), in collaboration with the National Center for Atmospheric Research (NCAR) and the National Oceanic and Atmospheric Administration (NOAA) are tasked by the Federal Aviation Administration (FAA) to provide aviation weather decision support tools for the air traffic management (ATM) community. To determine which weather products and data dissemination approaches will provide the greatest benefit in terms of increasing airspace capacity, MIT LL is performing ongoing marketing research analyses. The method consists of three primary steps (Ballentine 1994; Evans and Robinson 2005; Evans et al. 2003): 1) Study the system 2) Identify benefits 3) Prioritize opportunities In practice, the execution of these three steps is an iterative process. It is critical to understand how the air traffic system operates to assess the benefits of a weather product. For this reason many studies have been conducted by MIT LL where ATM users have been interviewed, use of decision support tools has been observed, and flight track data have been analyzed to extract the behavior of the pilot (Evans et al. 2003; Evans and Robinson 2005; Robinson et al. 2004; Rhoda et al. 2002; Allan et al. 2001; Bieringer et al. 1999; Rhoda and Pawlak 1999; Forman et al. 1999). These studies have provided valuable insight into how the NAS operates during weather impacts. Weather impacts on the air traffic system can be classified into three basic types: 1) Terminal impacts (≤ 5 nm) 2) En route impacts 3) Transition impacts (between Air Route Traffic Control Centers (ARTCC) sectors and terminal operations) Terminal impacts are those that occur in and around the airport, and are generally less than 5 nm from the runways. These impacts are small in dimension and occur at low altitude, and have been shown to be relatively insignificant to the overall delay problem (Evans et al. 2005). En route impacts occur within ARTCC's jet routes and sectors. These impacts can result in a route being totally or partially blocked and lead to a reduction in capacity. The en route impacts generally occur at high altitude. Transition impacts are those that occur within the zone between the ARTCC sectors and the terminal approaches. By understanding the system one can then identify elements or areas of opportunities that can be exploited to help solve the airspace capacity problem. Weber et al. (2005) identifies four key elements for maintaining capacity during convective weather events: 1) Forecasts of convective weather 2) Capacity models where weather is an input 3) Strategy tools for ATM with weather as an input 4) Airspace capacity enhancements Forman et al. (1999) studied the terminal impact problem and found that the ATM users required precipitation forecasts that were reliable, updated rapidly (5-6 minutes), had high resolution (1 km), short lead times (1-2 hours), and were issued with fine time steps (10-15 minutes). MIT LL used this information to refine its terminal convective weather forecast (TCWF) and received very positive feedback from ATM personnel (Hallowell et al. 1999). Subsequently, the precipitation forecast was extended out to 2-hourtime horizons and was provided to the traffic managers working in busy Midwest and Northeast ARTCCs as part of the Corridor Integrated Weather System (CIWS) (See Klingle-Wilson and Evans (2005) for a description of the CIWS product). However, it was quickly determined that the precipitation product alone was not sufficient for identifying usable en route airspace, since occasionally significant precipitation (≥ level 3)1 had relatively low storm tops (≤ 30 kft). Due to feedback from ATM users, MIT LL produced a high resolution (1 km) enhanced Echo Tops Mosaic weather product (Evans et al. 2003) that is used as a proxy for the cloud top height. Since the operational inception of the CIWS enhanced Echo Tops Mosaic product in August 2002, FAA and airline traffic managers have become acutely aware of the benefits of high-resolution storm top information for efficient en route air traffic control (ATC) operations. CIWS field use assessment campaigns in 2003 revealed significant benefits attributed to use of the Echo Tops Mosaic product (Robinson et al. 2004). During interviews, traffic managers explained that in the past, if an aircraft deviated around a storm in high-traffic airspace, jet routes were closed by default, since pilot behavior was the only easily assessable information available about three-dimensional storm structure. After the CIWS Echo Tops Mosaic was introduced, traffic managers were able to differentiate between isolated storm top concerns, which are easily handled by keeping routes open and absorbing occasional, local deviations, and significant high-topped storm events, which legitimately require route closures and reroutes. Post-event interviews in 2003 revealed that though FAA and airline users were very pleased with the availability and quality of the CIWS Echo Tops Mosaic product, they also needed to know both the past trend and predicted behavior of storm top heights. The CIWS Echo Tops Forecast (ETF) was introduced in May 2005 to meet some of the traffic management requests. This paper discusses the ETF product currently operational in CIWS. We will discuss the generation of the forecast algorithm and provide an initial assessment of the use of the ETF in the field.