Mari Ostendorf, Michiel Bacchiani, Y. Sagisaka, K. Paliwal
Speech Recognition System Design
using Automatically Learned
Non-Uniform Segmental Units
Abstract:This report describes a new approach to acoustic modeling that combines segmental acoustic
modeling with automatic learning techniques to determine the appropriate acoustic units. Since
the units are not constrained to be phonetic (or tied to any other linguistic unit) and can optionally
span word boundaries, an important problem is automatic determination of the lexicon, which
includes both the list of lexical entries and the "pronunciation" in terms of the segmental units.
The solution to these problems is outlined below, and some preliminary experimental results are
described.