Featured Alum: Eric Fosler-Lussier

Eric Fosler-Lussier is an assistant professor at Ohio State University. He worked at ICSI while he was an Electrical Engineering graduate student at UC Berkeley from 1994-1999, then as a postdoctoral researcher from 1999-2000. He left ICSI after being offered a position at Bell Labs/Lucent Technology, where he worked on development of dialog systems. When Lucent closed their speech group in 2002, he did a short research visit to Columbia University where he worked with our other featured alum Dan Ellis. He then secured an assistant professorship at Ohio State, where he has been since.

His research focus today is on the use of neural net detectors for phones (the basic units of pronunciation used in ASR, or automatic speech recognition) and features of phones that can be used to improve ASR accuracy. He is looking at the relationship between the phone predicted by a neural net detector versus the dictionary pronunicaton, and rethinking how ASR could work in light of this. This current work builds on ASR systems developed at ICSI: Nelson Morgan and Hervé Bourlard's original hybrid system, and the system that followed it, developed by Dan Ellis and Hynek Hermansky, a Gaussian based system using neural nets. Fosler-Lussier is now using feature detectors into conditional random fields (a slightly different statistical paradigm).

While at ICSI, Fosler-Lussier worked for the Speech Group on projects related to signal processing for ASR. In one project, he worked with Professor Nelson Morgan and Dr. Nikki Mirghafori on analyzing speaking rate. At the time, Mirghafori, now a staff scientist with the Speech Group, was writing her master's thesis on the effect of speaking rate on ASR. In an early project, they looked at the wall street journal corpus and the effect of speaking rate on word error rate (WER). Later, they used the Switchboard corpus, which was hand-transcribed at the word, syllable, and phoneme level, using the syllable count from the transcription to determine speaking rate. Morgan led the development of a signal processing method based on this work, and Fosler-Lussier did significant work on that system. They evaluated the output of signal processing versus the manual transcription of switchboard, and used it to predict reduced pronunciations (i.e., slurring syllables together) with regard to speaking rate. They also compared pronunciations from human transcripts to the dictionary pronunciations of the same words as well as to the pronunciation from the signal processing output, again with regard to syllable count and reductions. Eric also briefly worked on the SmartKom project (predecessor to the current SmartWeb project) at ICSI.

As a graduate student, he was a teaching assistant for an artificial intelligence class taught by ICSI emeritus board member Professor Jitendra Malik. For this course, Fosler-Lussier gave a lecture on Hidden Markov Models (HMMs), which was developed into an ICSI technical report, and has since been used as course material since then at several universities.