I am an Assistant Professor of Computer Science at Queens College (CUNY), and a member of the Doctoral Faculty of the Computer Science and Linguistics programs at the CUNY Graduate Center. I completed my Ph.D. at Columbia University in 2009. I lead the Speech Lab @ Queens College.
My research concerns Natural Language Processing, Spoken Language Processing, Prosody/Intonation and Machine Learning.
I am the main author and maintainer of AuToBI - an open source java toolkit for prosodic analysis in English.
Along with Julia Hirschberg and Stefanie Shattuck-Hufnagel, I am developing Reciprosody, a repository of prosodically annotated speech.
I collaborate part time at the IBM TJ Watson Research Lab, where I helped improve the speech synthesis quality for Watson, the Jeopardy! playing system.
At Interspeech 2011 in Florence, I gave a tutorial called "More than Words Can Say: Prosodic Analysis Techniques and Applications". The slides are available here.
Most recent blog post:
Based on my dissertation, "Automatic Detection and Classification of Prosodic Events", I developed AuToBI, a system to generate hypothesized ToBI labels.
Detecting Deception in Speech
People indicate that they are lying in different ways. While one
speaker might hesitate a lot while being deceptive, and very little
when being truthful, another speaker might demonstrate the opposite
We are investigating the lexical and acoustic correlates to deceptive
behavior. In this work, we are paying particular attention to two
qualities: cross-cultural differences in deceptive behavior, and 2)
the influence of a speaker's personality on their deceptive strategy.
Using Prosody to Improve Keyword Search in Low Resource Languages
The intonation of a spoken sentence can inform the prediction of the words that are being said. For example, questions contain different words than declarative statements. We are developing techniques to use intonational (or prosodic) information to improve keyword search or (spoken term detection) in low resource languages. Collaborators: IBM, Columbia University, New York University, RWTH Aachen, University of Cambridge, University of Southern California
Analysis of Political Speech
Political campaigns are centered around political speeches and debates. Even with big budget television ads, and netroots organizing, political rhetoric remains the most potent device for a politician to drive a message. We are examining the rhetorical qualities of political speech and debate performances to determine what makes a politician attractive to a population.
User Modeling based on Typing Behavior
People have different typing styles. Some people plan a sentence in
their head and then type it with no pausing or revision. Others type
in shorter spurts with many corrections. We are developing techniques
that explore the different physical, lexical and cognitive correlates
to pausing and revision in typing behavior. These correlates will
then be used to distinguish typists.
Analysis of Non-native Intonation
Intonation is typically the last aspect of spoken language acquired by non-native speakers. In this work, we examine the influence of Mandarin Chinese speakers' native language on their perception and production of intonation in American English.
V-Measure: Cluster Evaluation Measure
Evaluating the success of clustering algorithms is not a trivial task. In 2007, I proposed a conditional entropy-based measure called V-measure at EMNLP. This measure is described in V-Measure: A conditional entropy-based external cluster evaluation measure. The name comes from its similarity in calculation to F-measure, and its goal of evaluating cluster validity.
V-measure is the evaluation measure for the unsupervised evaluation of the SemEval 2010 Word Sense Induction and Disambiguation Task
Distributed here is a java package that implements V-Measure as well as a number of other cluster evaluation measures. Feel free to email with bug reports and feature requests