SUNFEST at Penn
VOICING DETECTION USING HOMOMORPHIC SPEECH PROCESSING AND FILTER BANK ANALYSIS
An effective method of detecting voicing has been desired as a means of preparing speech samples for analysis. Currently, the two methods used are Cepstrum and Filter Bank Analysis. However, neither one is always accurate in its detection of voicing. This report will discuss the importance of voicing detection and the processes of Cepstrum Analysis in detail. This report will also discuss the steps taken to combine the two processes to create a more accurate means of voicing detection.
TRANSDERMAL DNA DELIVERY USING AN ULTRASONIC TRANSDUCER
Avisors: Prof. J. Santiago-Aviles (EE) and Dr. P. Bloomfield (BE, Drexel)
Recent advances in transdermal delivery have allowed for the administration of higher molecular weight drugs noninvasively with the aid of ultrasound. Previously, only lower molecular weight drugs could be absorbed through the relatively impermeable skin as larger molecules cannot be readily taken up. The goal of this research was to extend the transdermal application of ultrasound beyond drug delivery to gene therapy with the administration of a DNA plasmid through the skin. This research was founded in the hopes that eventually a transdermal patch containing a transducer could deliver DNA to the bloodstream in a painless technique using low frequency ultrasound. Though this goal has not yet been reached, I had the opportunity to vary some of the many parameters involved in this research including frequency, pulse length and intensity among others. In vivo testing on mice allowed for the exploration of the advantages and constraints of transdermal DNA delivery using ultrasound on a complex biological system.
ANALYSIS OF EXERGY LOSSES IN COMBUSTION OF METHANE
Exergy is an important thermodynamic property which measures the useful work that can be produced by a substance, or the amount of work needed to complete a process. Unlike energy, exergy is not conserved; analysis of exergy losses provides information as to where the real inefficiencies in a system lie. In this paper, the exergy losses inherent in combustion are calculated and the effects of changing combustion parameters are determined. A simplified model of combustion is proposed that will allow a detailed breakdown of exergy losses to be made.
The pyroelectric anemometer (PA), a thermal gas flow meter device, offers some interesting advantages over other existing means of thermal mass flow measurements. A study of the effectiveness of PA systems on unstable (turbulent) flows was conducted. In this report we present results acquired over the summer using various types of PA systems. These systems were employed in the University of Pennsylvania 2" Suction Wind Tunnel (PSWT). Additional structures built to control the flow in the PSWT. Measurements were carried out with these structures. Results of the various measurements conducted this summer are presented in graphical form and some discussion presented. The data is available in disk format.
AUTOMATIC SPEECH RECOGNITION OF FRICATIVES ON THE CORTICON NEURAL COMPUTER USING FEATURE EXTRACTION
This project represents the synthesis of two previous independently done projects, the first being Christopher D. Donham's doctoral thesis from 1995 and the second being Ahmed Ali's research done in 1996. The goal is to implement Ali's simulated neural network for fricatives onto the Analog Neural Computer using the details and feature extractors from Donham's thesis. A feed-forward network is designed, consisting of level detectors followed by decision makers. Sounds from a speech database go to a bank of bandpass filters which separate the sounds in order of increasing frequency into 16 bands. The output of these bands is presented to the neural computer. The level detectors analyze the energy content per band, and send that information to the decision makers, which then decides whether or not a phoneme is the fricative it was designed to detect.
PATTERN RECOGNITION USING A NEURAL NETWORK
During the summer I have worked with pattern recognition using neural networks. The purpose of this project is to introduce me to the field of pattern recognition using neural networks. This research can be segmented into three parts: first, to acquire a background on neural networks, back propagation, pattern recognition and Matlab; second, to build a library of input vectors that are used to train and test the neural network; and third, to program, simulate, train and expand the neural network. In this paper, we describe the first part and the beginning of the second and third segments. We provide an introduction to neural networks and pattern recognition based on what we have learned. We also provide a quick introduction to Matlab and its use as a neural network simulator. Problems confronted during the development of the neural network will be discussed as well as the future directions of this research.
A SYSTEM FOR SPECTROELECTROCHEMICAL CHARACTERIZATION OF THIN FILMS FOR ELECTROCHEMICAL SENSORS
A device has been developed for the spectroelectrochemical characterization of thin films for electrochemical microsensors. This consists of a reaction chamber containing a transparent indium tin oxide (ITO) working electrode, a gold counter electrode, and a silver-silver chloride reference electrode. The reaction chamber is constructed using a standard 5 mL spectrophotometer cuvette, and allows for easy delivery of fluids while mounted in a spectrophotometer. Changes in the absorption spectrum of a K3Fe(CN)6 solution due to the redox processes taking place during cyclic voltammetry have been successfully observed using this device.
VOWEL RECOGNITON USING NEURAL NETWORKS
Despite the seeming ease with which humans extract meaningful information from the speech signal, achieving this with man-made systems has proven difficult. We hear and understand effort-lessly, yet how we do this remains a mystery. The methods which should be used to construct machines capable of recognizing speech are therefore unclear. Various methods have been and are being explored to allow machines to recognize speech. The approach employed here uses neural networks to examines speech at the phoneme level.