ICSI Research Review

Friday, October 14, 2011
1:45 - 5:00pm

Featured talks by ICSI research staff highlighting some of our latest results and new directions in computer science research. Talks will be given in the 6th floor lecture hall.

Agenda:

1:45 Introduction and Research Review

Prof. Nelson Morgan
ICSI Director and Speech Group Leader

2:00 "Increasing Robustness of Speech Processing with Cortically-Inspired Signal Processing"

Prof. Nelson Morgan
ICSI Director and Speech Group Leader

2:45 "Don't Multiply Lightly: Quantifying the Problems with Modeling Assumptions in Speech Recognition"

Dr. Steven Wegmann
Speech Group Member

3:45 Break

4:00 "Next Stop 100G: Scaling Network Security Monitoring to New Operational Challenges"

Dr. Robin Sommer
Networking Group Member

Abstracts:
"Increasing Robustness of Speech Processing with Cortically-Inspired Signal Processing"

Researchers have long been intrigued by the human capacity for speech processing, particularly under acoustic conditions that baffle our best machine implementations. Some basic characteristics of mammalian auditory systems have already been incorporated into routinely used signal processing methods. These methods, however, rely almost entirely on either psychoacoustic measurements or physiological measures from early in the auditory chain, e.g., at the auditory nerve. In the last decade, a number of laboratories, including ICSI, have begun investigating methods inspired by analyses of signals in the primary auditory cortex. These methods have already yielded reduced sensitivity to poor acoustics for speech recognition, speech/nonspeech discrimination, and speaker recognition. This talk will give some of the background for this work and conclude with recent results in all three areas.

"Don't Multiply Lightly: Quantifying the Problems with Modeling Assumptions in Speech Recognition"

In this talk, I will describe some recent results from project OUCH - Outing Unfortunate Characteristics of HMMs - which is a collaborative effort between myself, Dan Gillick (ICSI), and Larry Gillick (EnglishCentral Inc.). The central idea behind this project is that if we want to improve speech recognition performance through acoustic modeling, then we should first quantify how the current best acoustic model — the hidden Markov model (HMM) — fails to adequately model speech data and how these failures impact speech recognition accuracy. What we are undertaking is a form of diagnostic analysis that is an essential component of statistical modeling but, for various reasons, has been largely ignored in the field of speech recognition. In particular, we believe that previous attempts to improve upon the HMM have largely failed because this diagnostic information was not readily available.

I will describe how we use simulation and a novel sampling process to generate pseudo test data that deviate from the HMM in a controlled fashion. These processes allow us to generate pseudo data that, at one extreme, agree with all of the model's assumptions, and at the other extreme, deviate from the model in exactly the way real data does. In between, we precisely control the degree of data/model mismatch. By measuring recognition performance on this pseudo test data, we are able to quantify the effect of this controlled data/model residual on recognition accuracy. I conclude that the HMM independence assumptions present a far more serious problem than properly modeling the output distributions.

"Next Stop 100G: Scaling Network Security Monitoring to New Operational Challenges"

Network security research benefits significantly from interaction with real-world environments. Unfortunately, however, one often finds a striking gap between academia and operations, even though a close collaboration can be extremely fruitful for both sides. In this talk, I discuss how operational demands have driven significant enhancements to the open-source Bro network intrusion detection system, developed at ICSI. Bro is now a crucial part of network operations at many sites ranging from research to government to industry. In return, the resulting operational technology not only provides researchers with unprecedented capabilities for future studies, but its deployments have also led to new research directions that we are currently pursuing.