Skip to main content
[an error occurred while processing this directive]

[an error occurred while processing this directive]
HomePrograms >

2013 Neukom Scholars

Winners of the Neukom Scholars Program have been announced with awards of up to $1,000 per term for up to two terms.

The programs goal is to fund undergraduate students engaged in faculty-advised research in the development of novel computational techniques or the application of computational methods to problems in the Sciences, Social Sciences, Humanities, and the Arts.

Third and fourth year students are eligible to apply for these competitive grants. Winners for Summer, Fall and Winter terms are:

Endangered Sacramento Steelhead

Amy Van Scoyoc ‘13

Recent advances in genetics have provided biologists a new opportunity to examine patterns of parentage in natural populations. These parentage analysis studies are rapidly being conducted to examine fundamental questions in ecology, evolution, and conservation biology (Jones and Ardren 2003). However, progress on genetic tools has outpaced computational methods for parentage analysis. Here we propose to explore six distinct computational approaches to parentage analysis to determine which is most appropriate for analyzing a dataset from an endangered steelhead trout population in Battle Creek, California. Genetic data for this project has already been collected by the U.S. Fish and Wildlife Service (USFWS). A 2008 USFWS report was produced using a basic computational approach to parentage analysis. More work is needed to explore other computational methods that have recently been developed and may be more appropriate and statically powerful for this system. I will use funding from the Neukom Scholars Program to explore six computational methods of parentage analysis for this California steelhead trout dataset. Computational methods include: exclusion, categorical allocation, fractional allocation, parental reconstruction, full probability parentage analysis, and sibship reconstruction. Descriptions of each of these methods can be found in Jones et al. (2010). I will use freely available computational software available for download from university researchers. An example of one of the software packages can be found at:

This study will have important implications for understanding the conservation value of captive breeding programs (hatcheries) for aiding in the recovery of endangered steelhead populations in California. My project will build on a similar study system and conservation issue presented by Araki et al. (2007). In addition to the ecological and evolutionary findings, the results will have implications for water users and conservation. For example, can hatchery stocking alone maintain steelhead recovery in an area with massive demands on stream water, or are other interventions required? Applying computational methods is the only robust way to extract information from this data set.


Generalizing Learning Classifier System to High Dimensional Problems

Student: Gediminas Bertasius (Computer Science)
Faculty Advisor: Jason Moore (Genetics)

Identifying causal factors for various complex diseases is currently one of the most important and challenging tasks in the field of genetics. Epistasis (gene-by-gene) interaction and heterogeneity are the two main phenomena hampering accurate identification of predictive genetic markers. Epistasis refers to the phenomenon where expression of one gene is highly dependent on the presence of one or more other genes. Heterogeneity occurs when an individual set of features predict the same phenotype independently (latent class problem). While epistasis has received much attention recently, most of the methods to model heterogeneity are still ineffective. Due to highly complex non-linear relationships between the features and the noisy nature of the datasets, ordinary statistical techniques fail to produce accurate predictions.

In my project I intend to explore different methods of how to solve the problem outlined above, and the success of these methods. In order to avoid the situation where no rule from the population matches some specific instance, the greater number of wildcards has to be incorporated into the initial population of rules. However, as LCS explores more data instances, wildcards must be replaced with some specific “if-then” statements in order to increase the accuracy of the algorithm. Furthermore, in order to achieve a successful evolution of rules, one needs to choose a good initial subset of features and incorporate them into the rules so that the rules would match as many data instances as possible. Choosing an optimal subset of features may be done using feature selection algorithms (RelieF, genetic algorithms) or various feature selection criteria that are widely incorporated in the applications of decision trees. In this project, my goal will be to find an optimal way to extend LCS to higherdimensionality problems without loss in the accuracy of predictions.


Schizophrenia and Comorbid Substance Abuse Disorder

Charlotte Blumenfeld

Many of the estimated fifty-one million people who suffer from schizophrenia worldwide are also diagnosed with a comorbid substance abuse disorder. One of these disorders, alcoholism, can very often have devastating effects on the afflicted individual’s career prospects, financial stability, social relationships, and physical health, and can exacerbate the symptoms of schizophrenia. The best treatment currently available for suppression of the hallucinations, delusions, and disorganized thinking that are schizophrenia’s most debilitating symptoms is clozaril (often sold under the brand name Clozapine). Clozaril, a second generation antipsychotic medication, has serious drawbacks and was even removed from the market for a period of time due to its limited safety profile. While clozaril has been shown to be moderately successful in controlling auditory and visual hallucinations, it eventually causes fatalities in a significant percentage of patients due to its tendency to cause myocarditis, cardiomyopathy, pulmonary embolism, respiratory depression, and agranulocytosis. Patients medicated with clozaril also experience a host of other less serious but tough-to-manage and chronic side effects, such as tardive dyskinesia, seizures, drooling, and weight gain.

I plan to conduct a computational analysis of our data using Microsoft Excel and SPSS (the Statistical Program for the Social Sciences). Excel will be used for data storage, for basic analysis, and for graphical representations of the data. I will learn how to record and utilize macros so that I may synthesize and compare data sets across multiple runs of animals as well as across multiple treatment conditions. While the study is being conducting, Excel will also be used for basic statistical analyses in order to monitor changes in animal behavior over time. Once the study is complete, we will then use SPSS in order to do conduct more complex analyses. We will analyze changes in all of the behaviors described above, with a focus on alcohol intake and preference. To this end, we will perform repeated measures analysis of variance, which will allow us to test the effects of multiple variables including time and drug treatment on these behavioral measures. We will also perform post-hoc paired samples t-tests to investigate changes over time across groups and we will use Tukey tests to compare differences between groups. Once all data has been analyzed, I anticipate presenting this data and its importance in treating schizophrenia and co-occurring alcoholism as well as contributing to a publication on the same topic.


Language Revitalization Browser Game

Faculty Advisor: Professor James Stanford, Linguistics 
Student Leader: Bonita Längle, ENVS & AMES Double Major 
Student Programmer: Chandrasekar Ramesh, Computer Science Major

The purpose of this project is to create an open source browser game based on the format created by the Where are Your Keys1 (WAYK) Foundation that is designed to promote spoken fluency in any language in a short time period using a fun and interactive manner. The pilot language will be that of a local Native American tribe, the Abenaki, which has fewer than twenty fluent speakers left. This language was chosen specifically because one of the students, Bonita, is both a member of the tribe and a current student of learning the language. The long-term goal is to create a game that a person or tribe studying any endangered language can use to help promote fluency. 

This goal is to have the game created by the end of spring term 2013. This project will, with Neukom Insitute’s permission, start in the middle of winter term, to ensure it is complete by the above designated date, as Bonita is graduating this coming June. As such, we are requesting the following funding: $500 per student in winter term and $1000 per student in spring term, for a total of $3000 over a term and half of work. 


Ensemble Committees for Stock Return Prediction

Student: James Brofos, Advisor: Meifang Chu

Recent research by Huerta, Corbacho, & Elkan [3] on how nonlinear support vector machines can systematically identify stocks with high and low future returns focused on use of a single classier algorithm. I propose to extend this work by implementing a compilation of a number of classiers into an ensemble or committee model. Ensemble learners in financial forecasting are typically composed of k -Nearest-Neighbor clustering algorithms and neural network-based approaches [1, 2]. Very little prior work in financial prediction has seen the development of ensemble models that comprise constituent classiers beyond these ubiquitous ones.

Preliminary results suggest that the additional learning algorithms have the robustness and learning capability to add meaningful explanatory power to ensemble models. In particular, using data collected from Feldberg's research databases, a single relevance vector machine has been observed to provide predictive power comparable to (and in some cases exceeding) existing machine learners in the eld of stock prediction. I propose to explore these promising results by incorporating relevance vector machines, support vector machines, random forests, and k -Nearest-Neighbor clustering algorithms into a committee model. I hope to capitalize on the relative strengths of each algorithm to produce financial forecasting models of high accuracy.


 Effects of Anthropogenic Noise on Orthopteran Communication 

Robin Costello 
Faculty Advisor: Mark McPeek, Biology 

Acoustic signals of calling insects (order Orthoptera) are important for mate identification and territory defense. These calls, however, are susceptible to sound interference from both other species of singing insects and humans. While the call of one insect species has been documented to affect the temporal pattern and dominant frequency of the call of another insect species in the same habitat (Latimer and Broughton 1984, Greenfield 1988), little research has been done on the effect of anthropogenic (human generated) noise on Orthopteran communication. 

Funding from the Neukom Institute will provide air and ground transportation to their synthetic road site to allow me to meet the research team and record the calling insect species. From these recordings, I will extract acoustic characteristics including dominant frequency and temporal structure. I will use these characteristics to parameterize and refine auto-detection templates. After refining my templates, I will test them in a subset of recordings to determine the false positives and negatives for each species. Finally, I will use the detection templates to search several hundred hours of recordings from the project to determine whether calling activity changes when road noise is introduced, whether calling levels return to pre-manipulation levels during the manipulations, and whether effects differ among Orthopteran species. Auto-detection of animal sounds is a relatively new but rapidly expanding field. Because insect calls are relatively simple (when compared to bird calls), the application of auto-detection programs to scientific inquiry regarding insects is a good starting point for a technology that will have applications for conservation.


Using Phage Display to Discover Peptides for Targeting Disease Proteins

Alexandra Della Pia

Dr. Mark Spaller, Advisor (Dept. of Pharmacology & Toxicology, DHMC)

The proposed project will investigate protein-protein and protein-ligand interactions in order to enrich the study of compounds that could potentially be used as drugs or developing reagents to treat cancer and neurobiological disorders. Specifically, my work will aid in the discovery of new sequences and compounds via phage display to be used in further drug development.

In this project, my work will be the result of phage display, which is a simple and efficient means of testing for protein sequences by producing and amplifying many bacteriophage, or phage, which are viruses with a specific protein coat that infect bacteria and encode their own DNA. The phage that bind the best are then further examined for sequencing and this information can be used in eventual drug development.

The computational component of this project, SPECTRAFluor4, is important to determining which phage are the best binders. SPECTRAFluor4 is a computer program and device that measures the intensity of the absorbance signal as a result of the Enzyme Linked Immunosorbent Assay, or ELISA, procedure. The absorbance information provided by SPECTRAFluor4 is then analyzed in Microsoft Excel, where the absorbance intensity of the phage is divided by the intensity of the control. The higher the ratio of the absorbance, the more likely the protein sequence unique to the phage represents a good binder. Being able to measure absorbance singles with SPECTRAFluor4 and make calculations in Microsoft Excel are crucial components to knowing which phage should advance to DNA sequencing.


Modeling of Tokamak Edge Plasma Turbulent Heating and Transport

Manaure Francisquez 

Faculty advisor: Barrett Rogers (Physics & Astronomy)

The economic and environmental impact of raising energy costs is ubiquitous. Nearly nine out of ten Americans continue to drive to work despite rising fuel costs1. The food system is very fuel and transport dependent; 2007-2009 saw a ~300% increase in the price oil followed by a 60% increase in the price of food2. Transitioning to meet President Obama’s “all-of-the-above” pledge, which includes a million electric vehicles on the road in 2012, puts strains on the grid that can currently only be met with coal, oil and controversial natural gas. An upcoming option to meet this demand without crippling the environment is utility-scale nuclear fusion.

Fusion on Earth is achieved by magnetically confining a hot plasma3 (~108 degrees) in a toroidal vessel called a tokamak. These extraordinary temperatures allow tritium and deuterium to collide and recombine, yielding helium and an energetic neutron4. This energetic neutron is used to heat up water that powers a steam turbine in the same way current nuclear plants operate, with the key difference that no radioactive waste is produced or scarce fuels utilized.

This research aims to develop a fully consistent three-dimensional electromagnetic gyro fluid computational model for turbulent heating and transport at the edge of tokamak plasmas during H-mode scenarios. Gyro-fluid equations are obtained with a similar derivation to that of fluid equations, but the former provides more realistic instability growth rates that can be correlated to experimental data. However, a fine resolution is still necessary, hence the need for a parallel algorithm. This code will be 3D-MPI-decomposed allowing researchers to obtain results in practical computing times. The code will primarily be written in Fortran90, but will also make use of C, shell scripts and makefiles. IDL and python will also be used for data visualization. Class work and honors thesis research have provided me with extensive experience in all these languages and in codes for plasma applications. Reviewing turbulence theory and models throughout spring 2012 has also been instrumental in understanding the complexities of the code and building a clear algorithm design.


Spatially Explicit Biodiversity Modeling for Land Management Decision-Making

F. Charles Governali

The Natural Capital Project (NatCap) is a partnership among Stanford University's Woods Institute for the Environment, University of Minnesota's Institute on the Environment, The Nature Conservancy, and World Wildlife Fund. The NatCap team is developing and applying ArcGIS-based tools that combine biophysical and economic models to quantify expected return on investments in natural capital, in spatially explicit terms relevant to human well-being. NatCap’s publicly available ecosystem services modeling toolbox, InVEST (Integrated Valuation of Environmental Services and Tradeoffs), “enables decision-makers to quantify the importance of natural capital, to assess the tradeoffs associated with alternative choices, and to integrate conservation and human development” (http://www.naturalcapital The initial version of the InVEST toolbox is already being applied by national and state governments in Indonesia, Tanzania, China, Colombia, Ecuador, Canada, and the United States.

The objective of my research this summer will be to help NatCap improve InVEST’s spatially explicit biodiversity model. First, I will use groundwork laid by a recent meta-analysis (Gibson et al. 2011) to build a larger, global dataset to allow more robust testing of existing biodiversity models and fitting new model forms. Focusing specifically on tropical forest birds, one of the taxa most sensitive to land use change (Gibson et al. 2011), I will apply the data from a set of 138 published studies spanning 28 countries and 92 study landscapes to create a pan-tropical dataset of biodiversity indicators associated with a detailed land use matrix at each study site. Specifically, I will use ArcGIS software to georeference study sites and analyze existing data to create a matrix of land use classes within a buffer area around each site. The current dominant biodiversity models go beyond the simple species-area relationship of island biogeography theory and address the critical importance of the surrounding landscape matrix—not just undisturbed habitat patches—for determining biodiversity levels. Prominent models differ, however, in how the mosaic of variously disturbed habitats is accounted for mathematically in predictions of biodiversity loss by taxon. For example, Pereira and Daily’s (2006) model addresses the whole landscape fundamentally as a mosaic of habitat types (both disturbed and undisturbed) to which each species group has a particular biological “affinity”:

Koh and Ghazoul’s (2010) model starts from the ratio of remaining undisturbed habitat and adjusts the “z”th power to which this is raised by the sensitivity of species groups to the surrounding matrix: With the dataset that I will create, I will be able to compare the performance of these competing models in predicting the effects of land use change on biodiversity across a broad spatial scale, and I will be able to propose and implement improvements to these models. More accurate models of biodiversity change in response to anthropogenic factors are critical for effective long-term use of natural resources. Working in both MATLAB and ArcGIS, I will attempt to develop improved model forms and write code to apply these model forms for predicting biodiversity change in a spatially explicit fashion under a given set of alternative land use change scenarios. The general aim of my research work is to support enhancements in our ability to model the natural world and thus, hopefully, our ability to utilize it in a fashion that maintains human well-being on the long-term and conserves spectacular natural resources for future generations to enjoy.


Computational Methods on Minimal Pancyclic Graphs 

Sean Griffin 
Advisor: Sergi Elizalde, Math Department 

A graph is called pancyclic if for any number n between 3 and the number of vertices of the graph, there exists a cycle with that length. Let us call the minimum number of edges of all pancyclic graphs with n vertices m(n). A pancyclic graph with n vertices is thus called minimal if it has m(n) edges. Several bounds have been placed on m(n) for any n in general. However, exact values have only been found for up to n=22. The main goals of the project are to expand our information about this function m(n) by calculating m(n) for n greater than 22 and by placing tighter bounds on m(n) than have already been established through previous research. 

Java programming language will be used to construct graphs and count cycle lengths, checking if the graph is pancyclic by looking at all cycles in the graph. Although previous research has neglected to use the advantages of technology to calculate the values of m(n), I will introduce this technology in order to increase the known values of this function as part of my research objectives. For each n and small values of k, a program will be used to construct all possible pancyclic graphs on n vertices with k chords. The program will first start with a Hamiltonian graph, represented with the “circle.” It will then add k chords through the circle and then check if the graph is pancyclic. If the graph is pancyclic, the program has computed m(n) to be k+n (the number of edges in the “circle” plus the number of chords). Otherwise, it will keep going through all chord placements with k chords until it has exhausted all possibilities, then it will increase k by 1 and continue searching until a pancyclic graph is found (eventually one must be found because if the program gets to k=n, one of these chord placements will result in a pancyclic graph). 

In order to check if a Hamiltonian graph is pancyclic, the program will choose specific chords and then find the cycle lengths of all possible cycles going through only those chords and no other chords. By [1], we know that there are only at most 2 cycles that utilize an exact set of chords. Therefore, the speed of the computation will be determined by how well the program can eliminate repetitive cases when it places chords down. The success of the program will largely depend on how well it can analyze chord patterns and recognize symmetries between graphs. This is where most of the work will be in computing m(n) for large n.


Is NMR a Viable Platform for Measurement-Based Quantum Computing?

Alexander Meill
Advisor: Professor Chandrasekhar Ramanathan, Department of Physics and Astronomy 

Quantum Information Processing (QIP) is an exciting and quickly growing field of modern physics, one that could possibly take computing in a promising new direction.  While quantum information theory is advancing rapidly, the physical implementation is still relatively underdeveloped.
Currently the circuit model of quantum computing is the most commonly implemented model.  In this model, qubits are processed by sequences of unitary gates that are organized into “circuits”.  Each gate rotates the qubits, and at the end of the circuit, the qubits are measured.  In the measurement-based model, however, the data qubit is entangled to other ancillary qubits, and these ancillary qubits are then measured in such a way that the data qubit collapses to the desired state.  Though the two models each have their own advantages on various implementation platforms, one particular advantage of measurement-based computing is its resistance to, and easy correction of, errors.
I will begin by formulating an NMR experiment that implements a measurement-based processing procedure.  This experiment will then be simulated extensively, taking into account a full model of the experimental system.  To simulate these operations, I need to store 2N complex numbers, where N is the total number of qubits used.  With each operation in an algorithm requiring up to 3 ancillary qubits, the size of the data set grows quickly, making the simulation very computationally challenging.  After I design and simulate a working experiment I will determine its feasibility.  Factors like the number of qubits, run time, and accuracy will all determine how feasible the experiment is.
Successful implementation of measurement-based QIP in NMR would be an exciting development in the field of quantum information science.  While there are disadvantages inherent in NMR, experimental implementations always yield insight into the physical constraints involved in any computing procedure.  I will also examine how these experimental constraints could affect implementation on other platforms.

Re-Envisioning the Washington Monument Grounds 

Mingyu Kim 
Faculty advisor: Karolina Kawiaka in Studio Art 

In May 2012, Professor Karolina Kawiaka in Studio Art Department was announced the winner for the National Ideas Competition for the Washington Monument Grounds. The multi-stage competition attracted more than 500 participants from around the world, and an international public vote has resulted in the winner. The proposed design by Professor Kawiaka creates a national gathering place at the base of the Washington Monument and is called “The People’s Forum.” The design envisions a civic place that creates a public forum for gatherings and completes a vision of the Mall as America’s front lawn. 

I plan to help Professor Kawiaka develop 3-D computer model of her design by using the design software Rhino, developed by McNeel. Rhino provides the tools to accurately model and document designs ready for rendering, animation, engineering, and analysis. It can create, analyze, and translate NURBS curves, surfaces, and solids with no limits on complexity, degree, or size. Most importantly, Rhino offers uninhibited free-form modeling, extreme precision, unrestricted editing, and large projects organization, which means that as more time and effort is put into the creation and revision of the spaces and the structures the design will be more refined and therefore more likely to be realized. 

My responsibility will be to chiefly use Rhino as a design and an analysis tool: to test which complicated ideas will hold together, analyze which parts of the ideas are workable, and visualize a series of successful design elements. Then we will walk through the iterations for the National Mall in 3-D as well as 2-D, choose the best for the exhibition in Washington DC, and output to the laser-cutter and the rapid prototyper to create physical models. Through this process, “The People’s Forum,” which is not fully developed in 3-D, will gain much more realism and opportunity to breathe before the eyes of audience. I am confident that “The People’s Forum” will become more compelling through precisely and beautifully rendered physical model.


Hydrological Determinants of Mosquito Borne Disease: Benchmark Data Sets for Modeling

Ben Southworth '12
with Dorothy Wallace and Colleagues

Malaria is a major health problem in sub-Saharan Africa and the Philippines. Mathematical models are the key to understanding optimal delivery of interventions, such as introduction of bed nets, adulticide, and larvicide, especially regarding timing of adulticide and larvicide applications, often yielding non-intuitive results for vector borne disease. Numerical experiments can also test the effects of multiple simultaneous interventions; sthe closest one can come to a “controlled” experiment. Dynamic models given by systems of ODEs take as input the parameters controlling the spread of disease, producing predicted incidence of malaria cases as output. Variations in input such as bite rate, initial incidence of disease, prevalence of mosquitoes, etc., can therefore be explored to see their effects on malaria prevalence in the model. Inputs may be varied singly or in tandem. Causality is therefore clearer in these dynamic models than in data-driven correlations.

I will specifically be analyzing and simulating existing models using comparable rate constants, to determine key distinctions between them. Although there are many models which include mosquito dynamics, important input parameters for those dynamics are rarely available. Only two of the models found include larval stage in the mosquito dynamics. Even though mosquito dynamics drive malaria epidemics, data on these dynamics are completely missing. Thus not one single model of malaria dynamics has been validated against a complete data set that includes larval dynamics as an expected factor in disease. Many models which we have found have listed important parameters as “not available”, are assigned arbitrarily, or are drawn from a source that assigns the parameter arbitrarily.

The only model found where a reference is given for every parameter listed is Chitnis, using estimated equilibrium mosquito population from Western Kenya, bite rates from Khmer and New Guinea, mosquito-to-human transmission probabilities from a paper that estimates these from an earlier publication in 1974, and human-to-mosquito transmission probabilities from Africa. The parameters do not represent data collected in the same year, at the same location, or even necessarily for the same species of mosquito or strain of malaria. I will be running extensive simulations on each of these models, in an attempt to understand exactly how each parameter effects the output of malaria models, looking for consistency in trends across models. From here, I will be comparing these models with previous data sets, and new field data from Professor Wallace’s colleagues, in an attempt to understand the effect of assigning relatively arbitrary parameters, ignoring certain parameters all together, or cross contaminating, i.e. using values from previously done studies with different physical locations, time periods, and malaria strains. Primarily I will be using Matlab to perform these simulations and analysis.


The Neurophysiological Effects of Pten Point Mutations in Patients with Autism Spectrum Disorder

Tyrone DeSpenza Jr.
Faculty Advisor: Bryan Luikart, DHMC/DMS Neurobiology & Physiology

The research that will be conducted regards the effects of Pten mutations on the development and function of neurons. Pten is a tumor suppressor gene, and the Pten mutations that will be used are the same Pten mutations that are found in patients with ASD. ASD is a pervasive social disorder that is characterized by sensitivity towards sensory stimuli, inhibited social interactions, and repetitive actions. This behavior is also seen in mice that have been genetically engineered to knockout the gene Pten. Using stereotaxic surgery, engineered viruses expressing Pten point mutations will be injected into the dentate gyrus of mice. The anticipated outcome of this research is to observe similar neurophysiological abnormalities as seen in the Pten mutated neurons of ASD patients.

The key component of my work involves computational image analysis of neurons taken by confocal microscopy, and this allows for the characterization of the different Pten point mutations. I use Image J, a computer program, to analyze the soma sizes and spine densities of neurons. The soma sizes are determined by manually tracing the circumference of each neuron. Manually tracing a dendritic segment by hand and marking all dendritic protrusions as spines, dividing the number of protrusions by segment length using Excel, determines the spine densities. Excel is an essential component of my work because it also allows me to obtain the average soma sizes and dendritic densities of the point mutations. Further, computational analysis allows me to analyze data to achieve greater insight into the nature of the Pten point mutations.


Exploring Speech Recognition in Computer Assisted Language Learning

Crystal Ye (
Faculty Advisors: Sravana Reddy and Jim Stanford, Linguistics Dept.

With the rise of globalization, multilingualism has become increasingly important. Since the majority of Americans are raised as monolingual speakers, computer programs for second language (L2) learning are in high demand. We propose to explore improvements that can be made upon the speech recognition components of current L2 acquisition programs such as the popular Rosetta Stone®. We focus on one major issue.

Current renditions of Rosetta Stone successfully identify as correct or incorrect the L2 prosodies (patterns of stress or tone) produced by the speaker, but often fail to distinguish similar consonant and vowel phones. For example, English speakers often mistake the Spanish "v" to be the phones [v] or [b] rather than [], but the system does not identify such errors. Based on analyses of phonological transfer between the speaker’s indicated L1 and desired L2 languages, commonly confounded phonemes between the speaker’s L1 and L2, or L2 phonemes that do not exist in the speaker’s L1, we wish to guide speech recognition algorithms towards identifying the most common speaker errors.

While there has been previous research on this problem in the context of L2 tutoring systems, there are technical difficulties inherent in fine-grained phone recognition, especially with nonnative speakers – which we hypothesize is the reason that commercial programs like Rosetta Stone underperform on this task. We plan to conduct research on two fronts: exploring better computational models of fine-grained phone recognition for non-native speech, and incorporating relevant theoretical ideas from linguistics, particularly phonetics and second language acquisition.

Another contribution in this project will address the user interface for pronunciation correction in existing systems, which is restricted to a slowed-down version of a previously played sound file. It would be extremely beneficial, upon analysis of the speaker’s errors, to display a visual phonetic guide to the correct pronunciation, with the phones marked up so as to be understood by speakers with no training in linguistic annotation. The ideal final outcome of this project will either be a tutoring system for a single L1/L2 pair, with phonetic recognition that surpasses commercial systems, or an academic research paper detailing the results of various models for this task, with the results compared against published research, or both. We will begin with an existing open-source speech recognition engine like HTK or Sphinx, and explore several different models and algorithms to adapt it for improved L2 phone recognition.



Last Updated: 8/4/15