Minimum phone error and I-smoothing for improved discriminative training

Abstract

It is proposed that a novel type, called a rational type, be added to the set of existent types in high-level languages to enhance the accuracy of computation. Most fractions cannot be represented exactly in the real type or floating-point type that are currently available in high-level languages. In the proposed type, a real number is expressed as a ratio of two integers. It has already been established theoretically that computation using a rational format gives more accurate results for a class of problems than that using a floating-point format. A statistical analysis done by the authors showed that computation of values of natural logarithm, exponential, sine, square root, arctangent and the nth power of a variable were found to be more accurate virtually all the time when the rational format was used instead of the floating-point format.<>