Treffer: Research Statement

During my graduate school, I participated in the development of large vocabulary continuous speech recognition system, together with Aldebaro Klautau and our advisor Alon Orlitsky. We felt that there are many parts of the speech recognition system not well understood or justified, but because of the fine tuning they received over the decades of development had performance very hard to surpass. We examined those choices with modern machine learning tools and ideas, and with models that can offer strong mathematical performance guarantees. Consequently my research spanned many fields such as information theory, signal processing, statistical learning theory, and algorithms. The system started as a small vocabulary recognizer, as in [1] where we demonstrated a Java applet for voice web navigation. By June 2004, we will have a full large vocabulary decoder. That will not be the end of the road, but just the beginning as it would provide an excellent testbed for novel machine learning theories. The following paragraphs describe my prior research work as well as my current and future research interests. Universal compression and language modeling [2, 3, 4] In virtually all practical large vocabulary systems, the deployed language models are based on the n-gram approximation where a distribution is estimated for every word following a sequence of previous